Advanced Pattern Recognition Technologies With Applications To Bio Metrics

Advanced Pattern
Recognition Technologies
with Applications to
Biometrics
David Zhang
The Hong Kong Polytechnic University, Hong Kong
Fengxi Song
New Star Research Institute of Applied Technology , China
Yong Xu
ShenZhen Graduate School of Harbin Institute of Technology, China
Zhizhen Liang
City University of Hong Kong, Hong Kong
Hershey New York
Medi cal i nforMati on sci ence reference
Director of Editorial Content: Kristin Klinger
Director of Production: Jennifer Neidig
Managing Editor: Jamie Snavely
Assistant Managing Editor: Carole Coulson
Typesetter: Jennifer Henderson
Cover Design: Lisa Tosheff
Printed at: Yurchak Printing Inc.
Published in the United States of America by
Information Science Reference (an imprint of IGI Global)
701 E. Chocolate Avenue, Suite 200
Hershey PA 17033
Tel: 717-533-8845
Fax: 717-533-8661
E-mail: [email protected]
Web site: http://www.igi-global.com/reference
and in the United Kingdom by
Information Science Reference (an imprint of IGI Global)
3 Henrietta Street
Covent Garden
London WC2E 8LU
Tel: 44 20 7240 0856
Fax: 44 20 7379 0609
Web site: http://www.eurospanbookstore.com
Copyright 2009 by IGI Global. All rights reserved. No part of this publication may be reproduced, stored
or distributed in any form or by any means, electronic or mechanical, including photocopying, without written
permission from the publisher.
Product or company names used in this set are for identi.cation purposes only. Inclusion of the names of
the products or companies does not indicate a claim of ownership by IGI Global of the trademark or registered
trademark.
Library of Congress Cataloging-in-Publication Data
Advanced pattern recognition technologies with applications to biometrics / by David Zhang [et al.].
p. cm.
Includes bibliographical references and index.
Summary: This book focuses on two kinds of advanced biometric recognition technologies, biometric data
discrimination and multi-biometrics--Provided by publisher.
ISBN 978-1-60566-200-8 (hardcover) -- ISBN 978-1-60566-201-5 (ebook)
1. Biometric identifcation. I. Zhang, David, 1949-
TK7882.B56A35 2009
006.4--dc22
2008040444
British Cataloguing in Publication Data
A Cataloguing in Publication record for this book is available from the British Library.
All work contributed to this manuscript is new, previously-unpublished material. The views expressed in this
manuscript are those of the authors, but not necessarily of the publisher.
If a library purchased a print copy of this publication, please go to http://www.igi-global.com/agreement for
information on activating the library's complimentary electronic access to this publication.
Preface............................................................................................................... vii
Acknowledgment............................................................................................... xv
ChapterI
Overview.............................................................................................................. 1
Introduction ........................................................................................................... 1
Biometric Recognition Technologies .................................................................... 5
Main Problems in Biometric Recognition ............................................................. 8
Book Perspective ................................................................................................. 12
References ........................................................................................................... 16
SectionI
BiometricDiscriminantAnalysis
ChapterII
DiscriminantAnalysisforBiometricRecognition.......................................... 25
Linear Discriminant Analysis............................................ ..................................25
LDA for Solving the Small Sample Size Problems .............................................. 27
Organization of Section I .................................................................................... 28
References ........................................................................................................... 28
Table of Contents
Chapter III
Discriminant Criteria for Pattern Classi.cation ............................................ 30
Introduction.......................................................................................................... 30
Large Margin Linear Projection Classi.er ......................................................... 34
Minimum Norm Minimum Squared-Error Classifer........................................... 38
Maximum Scatter Difference Classifer............................................................... 43
Summary.............................................................................................................. 55
References............................................................................................................ 55
Chapter IV
Orthogonal Discriminant Analysis Methods....................................................58
Introduction.......................................................................................................... 58
Orthogonalized Fisher Discriminant................................................................... 63
Fisher Discriminant with Schur Decomposition.................................................. 65
Comparison of Orthogonal Discriminant Analysis Methods............................... 70
Summary.............................................................................................................. 75
References............................................................................................................ 76
Chapter V
Parameterized Discriminant Analysis Methods ............................................. 78
Parameterized Direct Linear Discriminant Analysis........................................... 78
Weighted LDA in the Range of Within-Class Scatter Matrix............................... 94
Summary............................................................................................................ 103
References.......................................................................................................... 103
Chapter VI
Two Novel Facial Feature Extraction Methods ............................................ 106
Multiple Maximum Scatter Difference............................................................... 106
Feature Extraction Based on Coeffcients of Variances......................................119
Summary............................................................................................................ 131
References.......................................................................................................... 132
Section II
Tensor Technology
Chapter VII
Tensor Space .................................................................................................... 135
Background........................................................................................................ 135
Basic Notations.................................................................................................. 136
Tensor Decomposition........................................................................................ 142
Tensor Rank........................................................................................................ 145
References.......................................................................................................... 148
ChapterVIII
TensorPrincipalComponentAnalysis.......................................................... 150
Introduction ....................................................................................................... 150
Basic Algorithms ............................................................................................... 152
Applications to Biometric Veri.cation .............................................................. 161
Summary and Discussion .................................................................................. 168
References ......................................................................................................... 170
ChapterIX
TensorLinearDiscriminantAnalysis............................................................ 172
Introduction ....................................................................................................... 172
Basic Algorithms ............................................................................................... 173
Applications to Biometric Verifcation .............................................................. 191
Summary and Discussion .................................................................................. 198
References ......................................................................................................... 199
ChapterX
TensorIndependentComponentAnalysisandTensorNon-Negative
Factorization.................................................................................................... 202
Introduction ....................................................................................................... 202
Tensor Independent Component Analysis ......................................................... 205
Tensor Non-Negative Factorization (NF) ......................................................... 213
Applications to Biometric Verifcation .............................................................. 220
Summary ........................................................................................................... 223
References ......................................................................................................... 223
ChapterXI
OtherTensorAnalysisandFurtherDirection.............................................. 226
Introduction ....................................................................................................... 226
Tensor-Based Classifers ................................................................................... 228
Other Tensor Subspace Analysis ....................................................................... 239
Summary ........................................................................................................... 249
References ......................................................................................................... 250
SectionIII
BiometricFusion
ChapterXII
FromSingleBiometricstoMulti-Biometrics................................................ 254
Introduction ....................................................................................................... 254
Biometric and Multi-Biometric Fusion: Defnition and Notation ..................... 257
Performance Evaluation of Biometric Techniques ........................................... 263
Research and Development of Multi-Biometrics .............................................. 265
References ......................................................................................................... 268
ChapterXIII
FeatureLevelFusion...................................................................................... 273
Introduction ....................................................................................................... 273
Schemes to Fuse Features at the Feature Level ................................................ 275
Face and Palm Print Fusion at the Feature Level ............................................ 276
Fusion of Multiple Feature Presentations ........................................................ 294
Comments .......................................................................................................... 301
References ......................................................................................................... 301
ChapterXIV
MatchingScoreLevelFusion......................................................................... 305
Introduction ....................................................................................................... 305
Matching Score Fusion Rules ........................................................................... 308
Normalization Procedures of Matching Scores ................................................ 314
Exemplifcation: Information Fusion of Face and Palmprint ........................... 318
Comments .......................................................................................................... 323
References ......................................................................................................... 323
ChapterXV
DecisionLevelFusion..................................................................................... 328
Introduction ....................................................................................................... 328
Rules and Methods of Decision Level Fusion ................................................... 330
Selecting Classifers Based on Correlations Between Classifers .................... 337
A Case Study of Group Decision-Based Face Recognition .............................. 338
Comments on Biometric Fusion at the Three Levels ........................................ 343
References ......................................................................................................... 345
ChapterXVI
BookSummary................................................................................................ 349
Content Summary .............................................................................................. 349
Method Applicability ......................................................................................... 351
Comments on Multi-Biometrics Development .................................................. 352
References ......................................................................................................... 357
Glossary........................................................................................................... 359
AbouttheAuthors........................................................................................... 364
Index................................................................................................................. 366
vii
Preface
With the increasing concerns on security breaches and transaction fraud, highly
reliable and convenient personal verifcation and identifcation technologies are more
and more requisite in our social activities and national services. Biometrics, which
use the distinctive physiological and behavioural characteristics to recognize the
identity of an individual, are gaining ever-growing popularity in an extensive array
of governmental, military, forensic, and commercial security applications.
The beginning of biometrics can be traced back to centuries ago, from when
fngerprint has been used for forensics. Automated biometrics, however, has only
40 years of history. In the early 1960s, the FBI (Federal Bureau of Investigation)
began to put more effort in developing automated fngerprint acquisition and
identifcation systems. With the advances in hardware, sensor, pattern recognition,
signal and image processing technologies, a number of biometric technologies, such
as face, iris, retina, voice, signature, hand geometry, keystroke, ear, and palm print
recognition, have been developed, and novel biometrics, such as dental, odor, and
skin refectance, have also been investigated to overcome some of the limitations
of current biometric recognition technologies.
Historically, the development of biometric technologies is originated from
different disciplines. For example, the beginning of fngerprint recognition research
is an interaction of forensics and pattern recognition. Voice recognition technology,
on the contrary, came from signal processing, and face recognition started from
computer vision. This multi-discipline characteristic, however, makes it very
challenging to establish an infrastructural theory framework for developing biometric
recognition technologies.
Generally, a biometric system can be regarded as a pattern recognition system,
where a feature set is frst extracted from the acquired data, and then compared with
the stored template set to make a decision on the identity of an individual. A biometric
system can be applied to two felds, verifcation and identifcation. In verifcation
mode, the decision is whether a person is who he claims to be? In identifcation
viii
mode, the decision is whose biometric data is this? A biometric system is thus
formalized into a two-class or multi-class pattern recognition system.
A biometric system usually includes four major modules: data acquisition,
feature extraction, matching, and system database, where feature extraction and
matching are two of the most challenging problems in biometric recognition research,
and have attracted researchers from different backgrounds: biometrics, computer
vision, pattern recognition, signal processing, and neural networks. In this book, we
focus on two advanced pattern recognition technologies for biometric recognition,
biometric data discrimination and multi-biometrics. Biometric data discrimination
technology, which extracts a set of discriminant features by using classical or
improved discriminant analysis approaches, is of course one kind of advanced
pattern recognition technology. Multi-biometrics, which integrates information from
multiple biometric traits to enhance the performance and reliability of the biometric
system, is another kind of advanced pattern recognition technology.
The book begins with the topic of biometric data discrimination technologies.
Discriminant analysis, which aims at dimensionality reduction while retaining
the statistical separation property between distinct classes, is a natural choice for
biometric feature extraction. From the late 1980s, many classical discriminant
analysis technologies are borrowed and applied to deal with biometric data or
features. Among them, principal component analysis (PCA, or K-L transform) and
Fisher linear discriminant analysis (LDA) turns out to be effective, in particular for
face representation and recognition. Other linear approaches, such as independent
component analysis (ICA), canonical correlation analysis (CCA), and partial least
squares (PLS), have been investigated and applied to biometric recognition. Recently,
non-linear projection analysis technology represented by kernel principal component
analysis (KPCA), kernel Fisher discriminant (KFD), and manifold learning, also
show great potential in dealing with biometric recognition problems.
Biometric data discriminant analysis is not a simple application of discriminant
analysis to biometric data. Biometric data is usually high dimensional and their
within-class variations should not be neglected, as the neglect will cause the seri-
ous performance degradation of classical discriminant analysis approaches. Various
improvements to discriminant analysis techniques have been proposed to make it
more suitable for biometric data. Now discriminant analysis has been widely applied
to face, ear, fngerprint, gait recognition, and multi-biometrics. Further, the present
demand for more reliable biometric recognition technologies is also contributing
to the development and improvement on linear/nonlinear discriminant analysis
technologies.
The form of biometric data and features is diverse. Biometric data mainly exists
in the following three forms: 1D waveform (e.g. voice, signature data), 2D images
(e.g. face images, fngerprints, palm prints) or image sequences (i.e., video), and
ix
3D geometric data (such as 3-D facial or hand geometric shapes). Because of the
diversity in biometric data and feature forms, it is diffcult to develop a universal
feature extraction technology which is capable to process all kinds of biometric
data.
To deal with the diversity in biometric data forms, a family of tensor discriminant
analysis technologies has been investigated. A tensor is a higher order generalization
of a vector or a matrix. In fact, a vector is a frst-order tensor and a matrix is a tensor
of order two. Furthermore speaking, tensors are multilinear mapping over a set of
vector spaces. If we have data in three or more dimensions, then we mean to treat
them as higher-order tensors. In this way, tensor technologies present a generalized
representation and analysis of biometric data discrimination technologies. Nowadays,
tensor principal component analysis, tensor discriminant analysis, tensor independent
component analysis, and other tensor analysis approaches have been successfully
applied to face, palm print, and gait recognition.
In advance of the application of biometric data discrimination technologies, we
should determine the appropriate representation of the biometric data. Generally,
biometric data discrimination technologies can be performed in either data space or
feature space. On the one hand, discrimination technologies can be used to derive
the discriminative features directly from the original biometric data. On the other
hand, a set of salient features are frst extract from biometric data, and discrimination
technologies are then performed in the feature space for the second feature extraction.
It should be noted that the feature space may be implicit or infnite dimensional
for kernel-based methods. In this way, biometric data discrimination technologies
provide a general means to integrate different kinds of features for effective biometric
feature extraction and matching.
In general, biometric data discrimination technologies share the following
characteristics: (1) high dimensionality, which makes direct classifcation in
original data space almost impossible; (2) difference in sample quality, where robust
discrimination technologies should be developed to address biometric data with
poor quality; (3) small sample size, where the data dimensionality is much higher
than the size of the training set, resulting in the singularity and poor estimation of
the scatter matrix. Recently, with the efforts of related researchers, progresses have
been achieved in solving these problems.
From the late 1990s, the authors of the book have been devoted to biometric data
discrimination research, and developed a series of novel and effective discriminant
criteria and discriminant algorithms, which cover both vector-based and tensor-based
discrimination technologies. The class of new methods includes: (1) Extensions of
Fishers discriminant criterion: we present three classifcation-oriented extensions
of Fishers discriminant criterion: large margin linear projection (LMLP), minimum
norm minimum squared-error, and maximum scatter difference. All these three
x
criteria are classifcation-oriented and designed to deal with the small sample size
problem; (2) Orthogonal discriminant analysis: we investigate two novel orthogonal
discriminant analysis methods: Orthogonalized Fisher discriminant (OFD) and Fisher
discriminant with Schur decomposition (FDS). Theoretical analysis and experimental
results indicate that both OFD and FDS are optimal solutions to multiple Fisher
discriminant criterion; (3) Parameterized discriminant analysis: we introduce three
parameterized linear discriminant analysis methods. The frst is parameterized direct
linear discriminant analysis. The second is weighted nullspace linear discriminant
analysis, and the third is weighted linear discriminant analysis in the range of
within-class scatter matrix; (4) New facial feature extraction methods: we present
two novel facial feature extraction methods, multiple maximum scatter difference
(MMSD) and discriminant based on coeffcients of variances (DCV). MMSD is
an extension of binary linear discriminant analysism and DCV is a generalization
of the null-space LDA; (5) Tensor PCA: we introduce the algorithms and discuss
the properties of a group of tensor PCA methods; (6) Tensor LDA: we investigate
the algorithms and discuss the properties of a group of tensor linear discriminant
analysis and tensor locality preserving projection methods.
The second topic of the book is multi-biometrics, which combines information
from multiple biometric traits to improve the accuracy and reliability of biometric
systems. Individual biometric system, the biometric system using a single biometric
characteristic, usually suffers from some inherent limitations and can not provide
satisfactory recognition performance. For example, manual workers with damaged
or dirty hands may not be able to provide high-quality fngerprint images, and thus
failure to enrol would happen for single fngerprint recognition system. Multi-bio-
metric systems, which integrate information from multiple biometric traits, provide
some effective means to enhance the performance and reliability of the biometric
system. In recent years, multi-biometric technologies have received considerable
interests in biometric recognition research. Several true multi-modal databases
have released for testing multi-biometric recognition algorithms. To facilitate
multi-biometric research, the (USA) National Institute of Science and Technology
(NIST) presents an open resource of Biometric Scores Set - Release 1 (BSSR1),
which includes true multi-model matching scores generated by face and fngerprint
recognition algorithms.
To combine information from individual biometric traits, there are three levels
of fusion strategies in general, feature level fusion, matching score level fusion, and
decision level fusion. In feature level fusion, the data obtained from each sensor
is used to compute a feature vector. As the feature extracted from one biometric
trait is independent of that extracted from the other, a new feature vector can be
constructed using the concatenation rule, the parallelity rule, or the competitive rule
for performing multi-biometric based personal authentication. It should be noted
xi
that the new feature vector may have a higher dimensionality than the original fea-
ture vector generated from each sensor, and feature reduction techniques may be
further employed to extract useful features from the set of the new feature vector. In
matching score level fusion, each subsystem using one biometric trait of the multi-
biometric system provides a matching score indicating the proximity of the feature
vector with the template vector. These scores can be combined to assert the veracity
of the claimed identity. A number of transformation-based, classifer-based, and
density-based score fusion methods have been used to combine scores of multiple
scores. In decision level fusion each sensor frst acquire one of multiple biometric
traits and the resulting feature vectors are individually classifed into the two deci-
sions - accept or reject the claimed identity. Then a scheme that exploits the known
decisions to make the fnal decision is used. So far, boolean conjunctions, weighted
decision methods, classical inference, Bayesian inference, DempsterShafer method,
and voting have been proposed to make the fnal recognition decision. In the feld of
multi-biometrics, a great number of studies of feature level fusion, matching score
level fusion and decision level fusion have been made. Though fusion of multi-bio-
metric are generally recognized as three classes as described above, in real-world
applications of multi-biometric system it is possible that the Fusion Process may
be simultaneously involved in different levels such as in both the matching score
level and the decision level.
In recent years, the authors and their collaborators have insisted on research on
multi-biometric technologies. Our investigations cover all the three categories of
multi-biometric technologies, which include: (1) Feature level fusion: we investigate
two novel feature level fusion methods, a pixel level fusion method to combine
face and palm print traits, and a feature level fusion method to combine phase and
orientation information of palm print images for effective personal recognition; (2)
Matching score level fusion: We present a practical application example to study
the effectiveness of matching score level fusion in face and palm print recognition;
(3) Decision level fusion: We introduce a group decision-making combination ap-
proach to combine decisions of multiple face recognition algorithms.
The book is organized into three main sections. As an overview of the book,
Chapter I describes the basic concepts necessary for a premier understanding of
biometric data discrimination and multi-biometric technologies. Section I explores
some advanced biometric data discrimination technologies developed for the small
sample size (SSS) problem, where we frst provide a brief introduction of linear
discriminant analysis and the SSS problem, and then describes our solutions to SSS
by developing extensions to Fishers discriminant criterion and novel improved
discriminant analysis approaches. Section II describes several tensor-based biometric
data discrimination technologies, including tensor principal component analysis,
tensor linear discriminant analysis, and tensor locality preserving projections.
xii
Other recently developed tensor approaches, such as tensor independent component
analysis, tensor non-negative matrix factorization, tensor canonical correlation
analysis, and tensor partial least squares, are also introduced in this section. Section
III deals with the second topic of the book, multi-biometrics. We frst introduce the
fundamental conception and categories of multi-biometrics technologies, and then
describe three kinds of multi-biometric technologies, feature level fusion, matching
score level fusion, and decision level fusion by providing several implementation
examples.
This book includes ffteen chapters. Chapter I briefy introduces two advanced
biometric recognition technologies, biometric data discrimination and multi-
biometrics, to enhance the recognition performance of biometric systems. In Section
1.1, we discuss the necessity, importance, and applications of biometric recognition
technology. A brief introduction of main biometric recognition technologies are then
presented in Section 1.2. In Section 1.3, we describe biometric data discrimination
and multi-biometric technologies. Section 1.4 outlines the history of related work
and highlights each chapter of this book.
Chapter II is a brief introduction to biometric data discriminant analysis
technologies. Section 2.1 describes two kinds of LDA methods: classifcation-ori-
ented LDA and feature extraction-oriented LDA. Section 2.2 discusses LDA for
the small-sample-size problem. Section 2.3 briefy introduces the organization of
Section I.
In Chapter III, we present three novel classifcation-oriented linear discriminant
criteria. The frst one is the large margin linear projection criterion, which makes
full use of the characteristic of the SSS problem. The second one is the minimum
norm minimum squared-error (MNMSE), which is a modifcation of the classical
minimum squared-error (MSE) discriminant criterion. The third one is the maximum
scatter difference, which is a modifcation of the Fishers discriminant criterion.
In Chapter IV, we frst give a brief introduction to Fishers linear discriminant,
Foley-Sammon discriminant, orthogonal component discriminant, and application
strategies for solving the SSS problem. We then present two novel orthogonal dis-
criminant analysis methods: one is orthogonalized Fisher discriminant; the other
is Fisher discriminant with Schur decomposition. At last, we compare the perfor-
mances of several main orthogonal discriminant analysis methods under various
SSS strategies.
Chapter V describes three kinds of weighted LDA methods. The first is
parameterized direct linear discriminant analysis. The second is weighted nullspace
linear discriminant analysis. The third is weighted LDA in the range space of within-
class scatter matrix. At last, we give a summery of the chapter.
Chapter VI introduces two novel facial feature extraction methods. The frst is
multiple maximum scatter difference (MMSD), which is extension of a binary linear
xiii
discriminant criterion, that is maximum scatter difference. The second is discriminant
analysis based on coeffcients of variances (DCV) which can be viewed as a gener-
alization of null space LDA. At last, we give a summery of the chapter.
Chapter VII gives the background for developing tensor-based discrimination
technologies. Section 7.2 introduces some basic notations in tensor space. Section
7.3 discusses several tensor decomposition methods. Section 7.4 introduces the
tensor rank.
Chapter VIII presents some variants of classical PCA, and discusses the proper-
ties of tensor principal component analysis. Section 8.1 gives the background and
development of tensor principal component analysis. Section 8.2 introduces tensor
principal component analysis. Section 8.3 discusses some potential applications
of tensor principal component analysis in biometric systems. We summarize this
chapter in Section 8.4.
In Chapter IX, classical LDA and its several variants are introduced. In some
sense, the variants of LDA can avoid the singularity problem and achieve compu-
tational effciency. Experimental results on biometric data show the usefulness of
LDA and its variants in some cases.
In Chapter X, we describe two subspace analysis technologies: tensor independent
component analysis and tensor non-negative matrix factorization, which can be used
in many felds like face recognition and other biometric systems. Section 10.1 gives
the background and development of two subspace analysis. Section 10.2 introduces
tensor independent component analysis. Section 10.3 gives tensor nonnegative fac-
torization. Section 10.4 discusses some potential applications of these two subspace
analysis in biometric systems. We summarize this chapter in Section 10.5.
Chapter XI deals with two tensor-based classifers, tensor canonical correlation
analysis and tensor partial least squares, which can be used in many felds such as
biometric systems. Section 11.1 briefy surveys the history in developing tensor-
based classifers. Section 11.2 introduces tensor-based classifers. Section 11.3 gives
tensor canonical correlation analysis and tensor partial least squares. We summarize
this chapter in Section 11.4.
Chapter XII describes the basic concepts of biometrics and motivation of multi-
biometrics. Section 12.2 provides the defnitions and notations of biometric and mul-
ti-biometric technologies. Section 12.3 presents two feature extraction approaches
for biometric images, where one of which is a linear feature extraction approach and
the other is a nonlinear feature extraction approach. Section 12.4 briefy indicates
some techniques associated with each kind of multi-biometric technologies.
Chapter XIII mainly presents the basic concepts and two examples of feature
level fusion methods. As the beginning of this chapter, Section 13.1 provides an
introduction to feature level fusion. Section 13.2 briefy surveys current feature
level fusion schemes. Section 13.3 shows a feature level fusion example that fuses
xiv
face and palm print. Section 13.4 shows a feature level fusion example that fuses
multiple feature presentations of a single biometric trait. Section 13.5 offers brief
comments.
In Chapter XIV, we aims at describing several basic aspects of matching score
level fusion. Section 14.1 provides a brief introduction of basic characteristics of
matching score fusion. Section 14.2 describes a number of matching score fusion
rules. Section 14.3 presents the normalization procedures of raw matching scores.
Section 14.4 provides several brief comments on matching score fusion.
Chapter XV deals with the decision fusion rules, the classifer selection approach,
and a case study of face recognition based on decision fusion, as well as a summary
of multi-biometric technologies. In a multi-biometric system, classifer selection
techniques may be associated with the decision fusion as follows: classifer selec-
tion is frst carried out to select a number of classifers from all classifer candidates.
Then the selected classifers make their own decisions and the decision fusion rule
is used to integrate the multiple decisions to produce the fnal decision. Section 15.1
provides an introduction to decision level fusion. Section 15.2 presents some simple
and popular decision fusion rules such as the AND, OR, RANDOM, and Voting rules,
as well as the weighted majority decision rule. Section 15.3 introduces a classifer
selection approach based on correlations between classifers. Section 15.4 presents
a case study of group decision-based face recognition. Section 15.5 offers several
comments on the three levels of multi-biometric technologies.
Chapter XVI aims at summarizing the book from a holistic viewpoint. Section
16.1 summarizes the contents of the book and indicates the relationship between
different chapters in each section. Section 16.2 reveals that how the methods and
algorithms described in different sections can be applied to different data forms
of biometric traits. Section 16.3 provides comments on the development of multi-
biometrics.
In summary, this book is a comprehensive introduction to theoretical analysis,
algorithms, and practical applications of two kinds of advanced biometric recognition
technologies, biometric data discrimination and multi-biometrics. It would serve as a
textbook or as a useful reference for graduate students and related researchers in the
felds of computer science, electrical engineering, systems science, and information
technology. Researchers and practitioners in industry and R&D laboratories working
on security system design, biometrics, computer vision, control, image processing,
and pattern recognition would also fnd much of interest in this book.
In the preparation of this book, David Zhang organizes the contents of the book
and is in charge of Chapters I, II, VII, XII and XVI. Fengxi Song handles Chapters
III through VI. Zhizhen Liang and Yong Xu write Chapters VII through XI and
Chapters XIII through XV, respectively. Finally, David Zhang looks through the
whole book and examines all chapters.
xv
The authors are full of gratitude to our co-workers and students for their persist-
ent support to our research. Our sincere thank goes to Prof. Zhaoqi Bian of Tsin-
ghua University, Beijing, Prof. Jingyu Yang of Najing University of Science and
Technology, Najing, China, and Prof. Pengfei Shi of Shanghai Jiaotong University,
Shanghai for their invaluable advice throughout this research. We would like to
thank our team members, Dr. Wangmeng Zuo, Prof. Jian Yang, Prof. Xiaoyuan
Jing, Dr. Guangming Lu, Prof. Kuanquan Wang, Dr. Xiangqian Wu, Dr. Jingqi Yan,
and Prof. Jie Zhou for their hard work and unstinting support. In fact, this book is
the collaborative result of their many contributions. We would also like to express
our gratitude to our research fellows, Feng Yue, Laura Liu, Dr. Ajay Kumar, Dr.
Lei Zhang, Dr. Hongzhi Zhang, Bo Huang, Denis Guo, and Qijun Zhao for their
invaluable help and support. Thanks are also due to Martin Kyle, for his help in the
preparation of this book. The fnancial support from the HKSAR Government, the
central fund from the Hong Kong Polytechnic University, the NFSC funds (Nos.
60332010, 60602038, and 60402018), and the 863 fund (No. 2006AA01Z193) in
China are of course also greatly appreciated. We owe a debt of thanks to Jan Travers
and Kristin Roth of IGI Global, for their valuable suggestions and keeping us on
schedule for the preparation and publishing of the book.
Acknowledgment
Overview 1
Copyright 2009, IGI Global, distributing in print or electronic forms without written permission of IGI Global
is prohibited.
Chapter I
Overview
Abstr Act
A biometric system can be regarded as a pattern recognition system. In this chapter,
we discuss two advanced pattern recognition technologies for biometric recogni-
tion, biometric data discrimination and multi-biometrics, to enhance the recognition
performance of biometric systems. In Section 1.1, we discuss the necessity, impor-
tance, and applications of biometric recognition technology. A brief introduction
of main biometric recognition technologies are presented in Section 1.2. In Section
1.3, we describe two advanced biometric recognition technologies, biometric data
discrimination and multi-biometric technologies. Section 1.4 outlines the history of
related work and highlights the content of each chapter of this book.
1.1 Introduct Ion
Reliable personal recognition techniques play a critical role in our everyday and
social activities. In access control, authorized users should be allowed for entrance
with high accuracy while unauthorized users should be denied. In welfare beneft
disbursement, people not only should verify whether the identity of a person is whom
he/she claimed to be, but also should avoid the occurrence that one person claims
to be another person to receive the welfare beneft twice (double dipping).
Traditionally, there are two categories of personal recognition approaches, to-
ken-based and knowledge-based (Miller, 1994). In the token-based approach, the
identity of a person is verifed according to what he/she has. Anyone possessed a
certain physical object (token), e.g., keys or ID cards, is authorized to receive the
2 Zhang, Song, Xu, & Liang
is prohibited.
associated service. The knowledge-based approaches authenticate the identity of
an individual according to what he/she knows. Any individuals with certain secret
knowledge, such as passwords and answers to questions, would receive the associ-
ated service. Both the token-based and the knowledge-based approaches, however,
have some inherent limitations. In the token-based approach, the token could be
stolen or lost. In the knowledge-based approach, the secret knowledge could be
guessed, forgotten, or shared.
Biometric recognition is an emerging personal recognition technology devel-
oped to overcome the inherent limitations of the traditional personal recognition
approaches (Jain, Bolle, & Pankanti, 1999a; Zhang, 2000, 2002, & 2004; Wayman,
2005; Bolle, 2004). The term biometrics, which comes from the Greek words bios
(life) and metrikos (measure), refers to a number of technologies to authenticate
persons using their physical traits such as fngerprints, iris, retina, speech, face and
palm print or behavior traits such as gait, handwritten signature and keystrokes. In
other words, biometric recognition recognizes the identity of an individual according
to who he/she is. Compared with the token-based and the knowledge-based methods,
biometric identifers cannot be easily forged, shared, forgotten, or lost, and thus can
provide better security, higher effciency, and increased user convenience.
Biometric recognition lays the foundation for an extensive array of highly secure
authentication and reliable personal verifcation (or identifcation) solutions. The
frst commercial biometric system, Identimat, was developed in 1970s, as part of
an employee time clock at Shearson Hamill, a Wall Street investment frm (Miller,
1994). It measured the shape of the hand and the lengths of the fngers. At the same
time, fngerprint-based automatic personal authentication systems were widely used
in law enforcement by the FBI and by US government departments. Subsequently,
advances in hardware such as faster processing power and greater memory capac-
ity made biometrics more feasible and effective. Since the 1990s, iris, retina, face,
voice, palm print, signature and DNA technologies have joined the biometric family
(Jain et al., 1999a; Zhang, 2000).
With the increasing demand for reliable and automatic solutions to security
systems, biometric recognition is becoming ever more widely deployed in many
commercial, government, and forensic applications. After the 911 terrorist attacks,
the interest in biometrics-based security solutions and applications increased dra-
matically, especially in the need to identify individuals in crowds. Some airlines
have implemented iris recognition technology in airplane control rooms to prevent
any entry by unauthorized persons. In 2004, all Australian international airports
implemented passports using face recognition technology for airline crews and
this will eventually became available to all Australian passport holders (Jain et
al., 1999a). Several governments are now using or will soon be using biometric
recognition technology. The U.S. INSPASS immigration card and the Hong Kong
Overview 3
is prohibited.
ID card, for example, both store biometric features for reliable and convenient
personal authentication.
Generally speaking, any situation that allows an interaction between human and
machine is capable of incorporating biometrics. Such situations may fall into a range
of application areas. Biometrics is currently being used in areas such as computer
desktops, networks, banking, immigration, law enforcement, telecommunication
networks and monitoring the time and attendance of staff. Governments across
the globe are tremendously involved in using and developing biometrics. National
identity schemes, voting registration and beneft entitlement programs involve the
management of millions of people and are rapidly incorporating biometric solu-
tions. Fraud is an ever-increasing problem and security is becoming a necessity in
many walks of life. Biometric applications can be simply categorized as follows
(Zhang, 2000):
LawEnforcement:The law enforcement community is perhaps the largest user
of biometrics. Police forces throughout the world use Automated Fingerprint Identi-
fcation System (AFIS) technology to process suspects, match fnger images and to
process accused individuals. A number of biometric vendors are earning signifcant
revenues in this area, primarily using AFIS and palm-based technologies.
Banking: Banks have been evaluating a range of biometric technologies for
many years. Automated Teller Machines (ATMs) and transactions at the point of
sale are particularly vulnerable to fraud and can be secured by biometrics. Other
emerging markets such as telephone banking and Internet banking must also be
totally secure for bank customers and bankers alike. A variety of biometric tech-
nologies are now striving to prove themselves throughout this range of diverse
market opportunities.
Computer Systems (also known as Logical Access Control): Biometric
technologies are proving to be more than capable of securing computer networks.
This market area has phenomenal potential, especially if the biometric industry can
migrate to large-scale Internet applications. As banking data, business intelligence,
credit card number, medical information and other personal data become the target
of attack, the opportunities for biometric vendors are rapidly escalating.
Physical Access: Schools, nuclear power stations, military facilities, theme
parks, hospitals, offces and supermarkets, across the globe employ biometrics to
minimize security threats. As security becomes more and more important for par-
ents, employers, governments and other groups - biometrics will be seen as a more
acceptable and therefore essential tool. The potential applications are infnite. Cars
is prohibited.
and houses, for example, the sanctuary of the ordinary citizen, are under constant
threat of theft. Biometrics - if appropriately priced and marketed - could offer the
perfect security solution.
Bene. t Systems: Beneft systems like welfare especially need biometrics to
struggle with fraud. Biometrics is well placed to capitalize on this phenomenal
market opportunity and vendors are building on the strong relationship currently
enjoyed with the benefts community.
Immigration:Terrorism, drug-running, illegal immigration and an increasing
throughput of legitimate travellers is putting a strain on immigration authorities
throughout the world. It is essential that these authorities can quickly and automati-
cally process law-abiding travellers and identify law-breakers. Biometric technolo-
gies are being employed in a number of diverse applications to make this possible.
The US Immigration and Naturalization Service is a major user and evaluator of
a number of state-of-the-art biometric systems. Systems are currently in place
throughout the US to automate the fow of legitimate travellers and deter illegal
immigrants. Elsewhere biometrics is capturing the imagination of countries such
as Australia, Bermuda, Germany, Malaysia and Taiwan.
NationalIdentity:Biometric technologies are beginning to assist governments
as they record population growth, identify citizens and prevent fraud from occur-
ring during local and national elections. Often this involves storing a biometric
template on a card that in turn acts as a national identity document. Finger scanning
is particularly strong in this area and schemes are already under way in Jamaica,
Lebanon, The Philippines and South Africa.
TelephoneSystems:Global communication has truly opened up over the past
decade, while telephone companies are under attack from fraud. Once again,
biometrics is being called upon to defend against this onslaught. Speaker ID is
obviously well suited to the telephone environment and is making in-roads into
these markets.
Time,AttendanceandMonitoring:Recording and monitoring the movement of
employees as they arrive at work, have breaks and leave for the day were traditionally
performed by time-card machines. Replacing the manual process with biometrics
prevents any abuses of the system and can be incorporated with time management
software to produce management accounting and personnel reports.
Overview 5
is prohibited.
1.2 bI ometr Ic r ecogn It Ion t echnolog Ies
A biometric system can be regarded as a pattern recognition system, where a feature
set is frst extracted from the acquired data, and then is compared with the stored
template set to make a decision on the identity of an individual. A biometric system
can be operated in two modes, biometric verifcation and biometric identifcation. In
biometric verifcation mode, the decision is whether a person is who he/she claims
to be? In biometric identifcation mode, the decision is whose biometric data is
this? Thus a biometric system can be formalized into a two-class or multi-class
pattern recognition system.
A biometric system usually includes four major modules: data acquisition, fea-
ture extraction, matching, and system database (Jain, Ross, & Prabhakar, 2004). In
the data acquisition module, the biometric data of an individual is acquired using
a capture sensor. In the feature extraction module, the acquired data is processed
to extract a set of discriminative features. In the matching module, the features
are compared with the stored template set to make a decision on the identity of an
individual. In the system database module, a database is built and maintained to
store the biometric templates of the enrolled users. Feature extraction and match-
ing are two of the most challenging problems in biometric recognition research,
and have attracted researchers from different backgrounds: biometrics, computer
vision, pattern recognition, signal processing, and neural networks.
Advances in sensor technology and increasing diverse demand of biometric
systems cause the persistent progress on developing novel acquisition sensors
and novel biometric technologies. Before 1980s, the offine ink-technique is the
dominant approach to acquire fngerprint images. Nowadays, a number of online
live-scan fngerprint sensors, e.g., optical, solid-state, and ultrasound, have been
designed for fngerprint acquisition.
Although research on the issues of common biometric technologies have drawn
considerable attention, and have been studied extensively over the last 25 years,
there are still some limitations to varieties of existing applications. For example,
some people have their fngerprints worn-away due to the hard work they do with
their hands and some people are born with unclear fngerprints. Face-based and
voice-based identifcation systems are less accurate and easier to be attacked us-
ing a mimic. Efforts geared towards improving the current personal identifcation
methods will continue, and meanwhile new biometric technologies are under inves-
tigation. Currently, the major biometric technologies involve face, fngerprint, iris,
palm print, signature, and voice recognition, as well as multi-biometric recognition
technologies (Zhang, 2002). The following provides a brief introduction of these
biometric traits:
is prohibited.
Fingerprint: Because the patterns of ridges and valleys on an individuals
fngertips are unique to that individual, fngerprints can be used for authenticating
personal identity. For decades, law enforcement has been classifying and determin-
ing identity by matching key points of ridge endings and bifurcations. Fingerprints
are so unique that even identical twins usually do not have the same fngerprint.
Iris: The patterns of the iris, the colored area that surrounds the pupil, are
thought to be unique. Iris patterns can be obtained through a video-based image
acquisition system. Iris scanning devices have been used in personal authentication
applications. It has been demonstrated that iris-based biometric system can work
with individuals without regard to ethnicity or nationality.
PalmPrint:Palm print, the inner surface of the palm, carries several kinds of
distinctive identifcation features for accurate and reliable personal recognition. Like
fngerprints, palm print have permanent discriminative features including patterns
of ridges and valleys, minutiae, and even pores in high resolution (>1000dpi) im-
ages. Except these quasi fngerprint features, palm print also carries other particular
distinctive features including principal lines and wrinkles. Using a high resolution
capture device, it is possible to extract all kinds of palm print features to construct
a high accurate biometric system. In the early stage, palm print recognition tech-
niques have been investigated to extract and match the singular points and minutia
points from high resolution palm print images. High resolution palm print scan-
ner, however, is expensive, and is time consuming to capture a palm print image,
which restricts the potential applications of online palm print recognition systems.
Subsequently, online capture device has been developed to collect real time low
resolution palm print image, and low resolution palm print recognition has gradu-
ally received considerable recent interest in biometric community (Zhang, Kong,
You, & Wong, 2003; Jain et al., 2004; Zhang, 2004).
Face:Images of a human face are highly suitable for use as a biometric trait
for personal authentication because they can be acquired non-intrusively, hands-
free, continuously, and usually in a way that is acceptable to most users (Zhao et
al., 2003). The authentication of a person by their facial image can be done in a
number of different ways, such as by capturing an image of the face in the visible
spectrum using an inexpensive camera or by using the infrared patterns of facial
heat emission. Facial recognition in visible light typically models key features from
the central portion of a facial image.
Signature:Signature authentication involves the dynamic analysis of a signa-
ture to authenticate a persons identity. A signature-based system will measure the
Overview 7
is prohibited.
speed, pressure, and angle used by the person when producing a signature. This
technology has potential applications in e-business, where signatures can be an
accepted method of personal authentication.
Speech:Speech-based personal authentication, which has a history of about four
decades, is regarded as a non-invasive biometric technology. Speech authentication
uses the acoustic features of speech, which have been found to be different between
individuals. These acoustic patterns refect both anatomic (e.g., size and shape of
the throat and mouth) and behavioral patterns (e.g., voice pitch, speaking style) of
an individual. The incorporation of learned patterns into the voice templates (the
latter called "voiceprints") has allowed speaker recognition to be recognized as
a "behavioral biometric". Speech-based personal authentication systems employ
three styles of spoken input: text-dependent, text-prompted and text-independent.
Most speech authentication applications use text-dependent input, which involves
selection and enrollment of one or more voice passwords. Text-prompted input is
used whenever there is concern about imposters.
Retina:The vascular confguration of the retina is supposed to be a character-
istic unique to each individual. The retina is regarded as one of the most secure
biometric traits. Usually the acquisition of a retinal image requires the cooperation
of the subject.
Gait:Gait, the peculiar way one walks, is a complex spatio-temporal biometric
trait. Note that gait is a behavioral trait and may not remain the same over a long
period of time, due to some factors such as changes in body weight. It is commonly
considered that a gait-based biometric system can be used in some low-security
applications. Gait authentication is also not intrusive and the acquisition of gait is
similar to acquiring a facial image. Usually a gait-based biometric system ana-
lyzes a video-sequence to obtain the gait trait and it is generally computationally
expensive.
HandandFingerGeometry:A system may measure geometrical characteristics
of either the hands or the fngers to perform personal authentication. These charac-
teristics include length, width, thickness and surface area of the hand. Requiring
only a small biometric sample of a few bytes is one interesting characteristic of
this kind of biometric technology. The biometric system based on hand and fnger
geometry has been used in physical access control in commercial and residential
applications, in time and attendance systems.
is prohibited.
DNA:Deoxyribonucleic acid (DNA), which can be denoted by a one-dimensional
code unique for every individual, is probably the most reliable biometric although
personal authentication using DNA will fail in distinguishing the identities of identi-
cal twins. A biometric system using DNA also suffers from other problems such as
privacy issues, possible contamination, and low-effciency in DNA matching.
Ear:There is evidence to show that the shape of the ear and the structure of the
cartilaginous tissue of the pinna are distinctive. As a result, the ear-based biometric
system can be used for authenticating personal identity.
Odor:Each object, including people, spreads an odor that is characteristic of
its chemical composition. This could be used for distinguishing various objects.
This would be done with an array of chemical sensors, each sensitive to a certain
group of compounds. However, deodorants and perfumes could compromise dis-
tinctiveness.
Multi-Biometrics:From an application standpoint, widespread deployment of
a user authentication solution requires support for an enterprises heterogeneous
environment. Often, this requires a multi-faceted approach to security, deploying
security solutions in combination. An authentication solution should seamlessly
extend the organizations existing security technologies. We are now interested
in understanding both how to build multi-biometric recognition systems and what
possible improvements these combinations can produce. Currently there are sev-
eral true multi-modal databases available for testing multi-biometric recognition
algorithms. The most important resource available may be the extended M2VTS
database, which is associated with the specifc Lausanne protocol for measuring
the performance of verifcation tasks. This database contains audio-visual material
from 295 subjects (Poh & Bengio, 2006). To facilitate multi-biometric research,
NIST presents an open resource of Biometric Scores Set - Release 1 (BSSR1),
which includes true multimodel matching scores generated by face and fngerprint
recognition algorithms (Grother & Tabassi, 2004).
1.3 mAIn Problems In bI ometr Ic r ecogn It Ion
To enhance the recognition performance of the biometric system, this section suggests
two advanced biometric recognition technologies, biometric data discrimination and
multi-biometric technologies. In biometric data discrimination, we frst introduce
the fundamental of biometric data discrimination, and then suggest using a family
Overview 9
is prohibited.
of tensor discriminant analysis to deal with the diversity in forms of biometric data.
In multi-biometrics, we introduce three categories of fusion strategies to enhance
the performance and reliability of the biometric system.
Besides recognition performance, security and privacy issues should also be taken
in account. In terms of security, there are many attacks, such as overplay, database
and brute-force attacks, on biometric applications. In terms of privacy, biometric
traits may carry additional sensitive personal information. For example, genetic
disorders might be inferred from the DNA data used for personal identifcation.
1.3.1 biometric data discrimination
Generally, biometric data mainly exists in the following three forms: 1D waveform
(e.g. voice, signature data), 2D images (e.g. face images, fngerprints, palm prints,
or image sequences, i.e., video), and 3D geometric data (such as 3-D facial or hand
geometric shapes). Since the diversity in biometric data and feature forms, it is
hardly diffcult to develop a universal recognition technology which is capable to
process all kinds of biometric data. Fortunately, recent progress in discriminant
analysis sheds some light on the possibility on this problem.
Discriminant analysis, with the goal of dimensionality reduction and of retain-
ing the statistical separation property between distinct classes, is a natural choice
for biometric recognition. With the development of biometrics and its applications,
many classical discriminant analysis technologies have been borrowed and applied
to deal with biometric data. Among them, principal component analysis (PCA, or
K-L transform) and Fisher linear discriminant analysis (LDA) have been very suc-
cessful, in particular for face image recognition. These methods have themselves
been greatly improved with respect to specifc biometric data analyses and appli-
cations. Recently, non-linear projection analysis technology represented by kernel
principal component analysis (KPCA) and kernel Fisher discriminant (KFD) has
also shown great potential for dealing with biometric recognition problems. In
summary, discrimination technologies play an important role in the implementa-
tion of biometric systems. They provide methodologies for automated personal
identifcation or verifcation. In turn, the applications in biometrics also facilitate
the development of discrimination methodologies and technologies, making dis-
crimination algorithms more suitable for image feature extraction and recognition.
Currently discriminant analysis has been widely applied to face, ear, fngerprint,
gait recognition, and even multi-modal biometrics. Further, the increasing demand
for reliable and convenient biometric system is also contributing to the development
and improvement of linear/nonlinear discriminant analysis techniques.
A tensor is a higher order generalization of a vector or a matrix. In fact, a vector
is a frst-order tensor and a matrix is a tensor of order two. Furthermore speaking,
is prohibited.
tensors are multilinear mapping over a set of vector spaces. If we have data in
three or more dimensions, then we mean to deal with a higher-order tensor. Tensor
presents a generalized representation of biometric data. To deal with the diversity
in biometric data forms, a family of tensor discriminant analysis technologies
have been investigated. Nowadays, tensor principal compoenent analysis, tensor
discriminant analysis, tensor independent component analysis, and other tensor
analysis approaches have been successfully applied to face, palm print, and gait
recognition.
Biometric data discrimination technologies can be briefy defned as automated
methods of feature extraction and recognition based on given biometric data. It
should be stressed that the biometric data discrimination technologies are not the
simple application of classical discrimination techniques to biometrics, but are
in fact improved or reformed discrimination techniques that are more suitable
for biometric applications, for example by having a more powerful recognition
performance or by being computationally more effcient for feature extraction or
classifcation. In other words, the biometric data discrimination technologies are
designed for extracting features from biometrics data, which are characteristically
high-dimensional, large scale, and offer only a small sample size. The following
explains these characteristics more fully.
High Dimensionality: In biometric recognition, high dimensional data usu-
ally are expected to be more powerful. The high-dimensionality of biometric data,
however, would make direct classifcation (e.g. the so-called correlation method that
uses a nearest neighbour classifer) in original data space almost impossible, frstly
because the similarity (distance) calculation is very computationally expensive,
secondly because it demands large amounts of storage. High dimensionality makes
it necessary to use a dimension reduction technique prior to recognition.
Large Scale: Real-world biometric applications are often large-scale, which
means biometric systems should be operated in large population databases. Typical
examples of this would include welfare-disbursement, national ID cards, border
control, voter ID cards, drivers licenses, criminal investigation, corpse identifca-
tion, parenthood determination, and the identifcation of missing children. Given
an input biometric sample, a large-scale biometric identifcation system determines
whether the pattern is associated with any of a large number (e.g., millions) of en-
rolled identities. These large-scale biometric applications require high-quality and
very generalizable biometric data discrimination technologies.
SampleQuality:Biometric systems automatically capture, detect and recogniz-
ing biometric image, making it inevitable that biometric data will sometimes be
Overview 11
is prohibited.
noisy or partially corrupted. The capture and communication of biometric data itself
may introduce noise; some accessories will cause the partial corruption of biometric
data, for example a scarf may occlude a facial image. Because all these factors are
inevitable, the development of biometric system should always address the robust
feature extraction and recognition of noisy or partially corrupted biometric data.
SmallSampleSize:Unlike, for example, optical character recognition (OCR)
problems, the training samples per class that are available in real-world biometric
recognition problems are always very limited. Indeed, there may be only one sample
available for each individual. Combined with high-dimensionality, small sample
size creates the so-called small sample size (or under-sampled) problems. In these
problems, the within-class scatter matrix is always singular because the training
sample size is generally less than the space dimension. As a result, the classical
LDA algorithm becomes infeasible in image vector space.
1.3.2 multi-biometrics
Verifcation or identifcation accuracy is always the frst-of-all objective for biomet-
ric systems. Unibiometric system, the biometric system using a single biometric
characteristic, usually suffers from some limitations and can not provide satisfac-
tory recognition performance. For example, manual workers with damaged or dirty
hands may not be able to provide high-quality fngerprint images, and thus failure
to enrol would happen for single fngerprint recognition system.
Multi-biometric systems, which integrate information from multiple biometric
traits, provide some effective means to enhance the performance and reliability of
the biometric system. To combine information from individual biometric traits,
there are three categories of fusion strategies, feature level fusion, matching score
level fusion, and decision level fusion. In feature level fusion, the data obtained
from each sensor is used to compute a feature vector. As the feature extracted from
one biometric trait is independent of that extracted from the other, it is reasonable
to concatenate the two vectors into a single new vector for performing multi-bio-
metric based personal authentication. Note that the new feature vector now has a
higher dimensionality than the original feature vector generated from each sensor.
Feature reduction techniques may be employed to extract useful features from the
set of the new feature vector. In matching score level fusion, each subsystem using
one biometric trait of the multi-biometric system provides a matching score indi-
cating the proximity of the feature vector with the template vector. These scores
can be combined to assert the veracity of the claimed identity. In decision level
fusion each sensor frst acquire one of multiple biometric traits and the resulting
is prohibited.
feature vectors are individually classifed into the two decisions---accept or reject
the claimed identity. Then a scheme that exploits the known decisions to make the
fnal decision is used. In the feld of multi-biometrics, a great number of studies
of feature level fusion, matching score level fusion and decision level fusion have
been made. Though fusion of multi-biometrics are generally recognized as three
classes as described above, in real-world applications of multi-modal biometric it
is possible that the Fusion Process may be simultaneously involved in different
levels such as in both the matching score level and the decision level.
1.4 book Pers Pect Ive
In this book we will systematically introduce readers to two categories of advanced
biometric recognition technologies, biometric data discrimination and multi-biomet-
rics. This book addresses fundamental concerns of relevance to both researchers and
practitioners using biometric data discrimination and multi-biometric technologies.
The materials in the book are the product of many years of research on the part of
the authors and present the authors recent academic achievements made in the
feld. For the sake of completeness, and readers may rest assured that wherever
necessary this book also addresses the relevant work of other authors.
Recent decades have witnessed the development and prosperity of biometric data
discrimination technologies. Various unsupervised/supervised, linear/nonlinear,
vector/tensor discrimination technologies have been investigated and successfully
applied to biometric recognition. At the beginning, linear unsupervised method,
principal component analysis (PCA), was used to extract the holistic feature vectors
for facial image representation and recognition (Sirovich & Kirby, 1987; Kirby &
Sirovich, 1990; Turk & Pentland, 1991a & 1991b). Since then, PCA has been widely
investigated and has become one of the most successful approaches to face rec-
ognition (Pentland, Moghaddam, & Starner, 1994; Pentland, 2000; Zhao & Yang,
1999; Moghaddam, 2002; Zhang, 2002; Kim, H. C., Kim, D., Bang, & Lee, 2004)
and palm print recognition (Lu, Plataniotis, & Venetsanopoulos, 2003b). Other
popular unsupervised methods, such as independent component analysis (ICA) and
non-negative matrix factorization (NMF), have been applied to biometric recogni-
tion (Bartlett et al., 2002; Yuen & Lai, 2002; Liu & Wechsler, 2003; Draper, Baek,
Bartlett, & Beveridge, 2003; Petridis & Perantonis, 2004).
Since the unsupervised methods do not utilize the class label information in the
training stage, it is generally believed that the supervised methods are more effective
in dealing with recognition problems. Fisher linear discriminant analysis (LDA),
which aims to fnd a set of the optimal discriminant projection vectors that map
the original data into a low-dimensional feature space, is then gaining popularity
Overview 13
is prohibited.
in biometric recognition research. In 1986, Fisher linear discriminant analysis was
frst applied to image classifcation (Tian, Barbero, Gu, & Lee, 1986). Further, LDA
was applied to face recognition, and subsequently was developed into one of the
most famous face recognition approaches, Fisherfaces (Liu, K. Cheng, Yang, &
Liu, X. 1992; Swets & Weng, 1996; Belhumeur, Hespanha, & Kriengman, 1997). In
biometric recognition, the data dimensionality is much higher than the size of the
training set, leading to the well-known small sample size (SSS) problem. Currently
there are two popular strategies to solve the SSS problem, the transform-based and
the algorithm-based (Yang & Yang, 2003; Jian, Yang, Hu, & Lou, 2001; Chen, Liao,
Lin, Kao, & Yu, 2000; Yu & Yang, 2001; Lu et al., 2003a; Liu & Wechsler, 2000 &
2001; Zhao et al., 1998; Loog, Duin, & Haeb-Umbach, 2001; Duin & Loog, 2004;
Ye, 2004; Howland & Park, 2004). The transform-based strategy frst reduces the
dimensions of the original image data and then uses LDA for feature extraction.
Typical transform-based methods include PCA+LDA and uncorrected LDA. The
algorithm-based strategy fnds an algorithm for LDA that can circumvent the SSS
problem. Some representative algorithm-based methods can avoid the SSS problem,
but most algorithm-based methods are computationally expensive or lose parts of
important discriminatory information.
Biometric recognition usually is highly complex and can not be regarded as a
linear problem. In the last few years, a class of nonlinear discriminant analysis tech-
niques named as kernel-based discriminant analysis has been widely investigated
for biometric data discrimination. Kernel principal component analysis (KPCA)
and kernel Fisher discriminant (KFD) are two of the most representative nonlinear
methods and have received considerable interests in the felds of biometrics, pat-
tern recognition, and machine learning. By far, a number of kernel-methods, such
as KPCA, KFD, complete kernel Fisher discriminant (CKFD), and kernel direct
discriminant analysis (KDDA), have been developed from biometric recognition
(Schlkopf et al., 1998; Mika Rtsch, Schlkopf, Smola, Weston, & Mller,1999a &
1999b; Baudat & Anouar, 2000; Roth & Steinhage, 2000; Mika, Rtsch, & Mller,
2001a & 2001b; Mika et al., 2003; Yang, 2002; Lu et al., 2003b; Xu, Zhang, & Li,
2001; Billings & Lee, 2002; Gestel, Suykens, Lanckriet, Lambrechts, De Moor,
& Vanderwalle, 2002; Cawley & Talbot, 2003; Lawrence & Schlkopf, 2001; Liu,
2004; Yang, Zhang, & Lu, 2004a & 2004b; Xu, Yang, J. Y, & Yang, J., 2004; Yang,
Zhang, Yang, Zhong, & Frangi, 2005). Most recently, manifold learning methods,
such as isometric feature mapping (ISOMAP), locally linear embedding (LLE),
and Laplacian eigenmaps, have also shown great potential in biometric recognition
(Tenenbaum, 2000; Roweis & Saul, 2000; Belkin & Niyogi, 2002).
As a generalization of vector-based methods, a number of tensor discrimination
technologies have been proposed. The beginning of tensor discrimination technol-
ogy can be traced back to 1993, where a 2D image matrix based algebraic feature
is prohibited.
extraction method is proposed for image recognition (Liu, Cheng, & Yang, 1993).
As a new development of the 2D image matrix based straightforward projection
technique, a two-dimensional PCA (2DPCA) and uncorrelated image projection
analysis were suggested for face representation and recognition (Yang, Zhang,
Frangi, & Yang, 2004c; Yang. J, Yang, J. Y., Frangi, A. F., & Zhang, 2003b). To
reduce the computational cost of 2DPCA, researchers have developed several
BDPCA and generalized low rank approximation of matrices (GLRAM) approaches
(Zuo, Wang, & Zhang, 2005; Ye, 2004; Liang & Shi 2005; Liang, Zhang, & Shi,
2007). Motivated by multilinear generalization of singular vector decomposition,
a number of alterative supervised and unsupervised tensor analysis methods have
been proposed for image or image sequence feature extraction (Lathauwer, Moor,
& Vndewalle, 2000; Vasilescu, & Terzopoulos 2003; Tao, Li, Hu, Maybank, & Wu,
2005; Yan, Xu, Yang, Zhang, Tang, & Zhang, 2007).
Over the last several years, we have been devoted to biometric data discrimination
research both in theory and in practice. A series of novel and effective technolo-
gies has been developed in the context of supervised and unsupervised statistical
learning concepts. The class of new methods includes:
ExtensionsofFishersDiscriminantCriterion: We present three classifca-
tion-oriented extensions of Fishers discriminant criterion: large margin linear
projection (LMLP), minimum norm minimum squared-error, and maximum
scatter difference (MSD). All these three criteria are designed to deal with
the small sample size problem;
Orthogonal Discriminant Analysis: We investigate two novel orthogonal
discriminant analysis methods: orthogonalized Fisher discriminant (OFD) and
Fisher discriminant with Schur decomposition (FDS). Theoretical analysis and
experimental studies showed that OFD and FDS are all optimal solutions to
multiple Fisher discriminant Criterion;
ParameterizedDiscriminantAnalysis:We introduce three parameterized
linear discriminant analysis methods, parameterized direct linear discriminant
analysis, weighted nullspace linear discriminant analysis, and weighted linear
discriminant analysis in the range of within-class scatter matrix;
Newfacialfeatureextractionmethods:We present two novel facial feature
extraction methods, multiple maximum scatter difference (MMSD) and dDis-
criminant based on coeffcients of variances (DCV);
Tensor PCA: We introduce the algorithms and discuss the properties of a
group of tensor PCA methods;
TensorLDA:We investigate the algorithms and discuss the properties of a
group of tensor LDA methods.
Overview 15
is prohibited.
Multi-biometric system is designed to overcome the limitations of any single
biometric systems by fusing information from multiple biometric traits. The fusion
can be implemented in either of three levels, feature level, matching score level,
and decision level. In feature level fusion, a new feature vector is constructed us-
ing the concatenation rule (Ross & Govindarajan, 2005), the parallel rule (Yang et
al., 2003a; Yang, J., & Yang, J. Y., 2002), or the competitive rule (Kong, Zhang, &
Kamel, 2006). In matching score level fusion, a number of transformation-based
(Jain, Nandakumar, & Ross, 2005; Zuo, Wang, Zhang, D., & Zhang, H., 2007), clas-
sifer-based (Brunelli & Falavigna, 1995; Jain, Prabhakar, & Chen, 1999b; Fierrez-
Aguilar, Ortega-Garcia, Gonzalez-Rodriguez, & Bigun, 2005), and density-based
(Ulery, Hicklin, Watson, Fellner, & Hallinan, 2006; Nandakumar, Chen, Jain, &
Dass, 2006) score fusion methods have been used to combine scores of multiple
scores. In decision level fusion, boolean conjunctions, weighted decision methods,
classical inference, Bayesian inference, DempsterShafer method, and voting have
been proposed to make the fnal recognition decision (Gokberk, Salah, & Akarun,
2003; Jing, Zhang, D., & Yang, 2003).
In the last part of this book, we summarize our recent research on multi-biometrics.
Our investigation covers all the three categories of multi-biometric technologies. The
class of new methods includes:
Featurelevelfusion:We investigate two novel feature level fusion methods, a
pixel level fusion method to combine face and palm print traits, and a feature
level fusion method to combine phase and orientation information of palm
print images;
Matchingscorelevelfusion:We present an example to illustrate the effec-
tiveness of matching score level fusion in face and palm print recognition;
Decisionlevelfusion: We introduce a group decision-making combination
approach to combine decisions of multiple face recognition algorithms.
This book is organized into three main parts. Chapter I frst describes the basic
concepts necessary for a premier understanding of biometric data discrimination and
multi-biometrics. Section I explores some advanced biometric data discrimination
technologies for the small sample size problem (SSS). Chapter II provides a brief
introduction of LDA and SSS. Chapter III describes our researches on the pattern
classifcation aspect of LDA for SSS problems. Chapter IV-VI present our studies
on the feature extraction aspect of LDA for SSS problems. Specially, Chapter IV
focuses on orthogonal discriminant analysis methods. Chapter V discusses pa-
rameterized discriminant analysis methods. Chapter VI describes two novel facial
feature extraction methods.
is prohibited.
Section II focuses on tensor-based biometric data discrimination technologies.
Chapter VII describes the background materials for developing tensor-based dis-
crimination technologies. In Chapter VIII, we develop some variants of classical
PCA and discuss the properties of tensor PCA as a generalization of PCA. Chapter
IX proposes two novel tensor discrimination technologies, two-dimensional LDA
and two dimensional locality preserving projection. Chapter X describes tensor
independent component analysis and tensor non-negative factorization, and Chap-
ter XI introduces some other tensor discrimination technologies, such as tensor
canonical correlation analysis and tensor partial least squares.
Section III states several multi-biometric technologies. Chapter XII introduces
the fundamental conception and categories of multi-biometric technologies. In
Chapter XIII, we develop two novel feature level fusion methods, a pixel level fu-
sion method and a phase and orientation information fusion method. Chapter XIV
dedicates to matching score level fusion. Chapter XV introduces decision level
fusion and proposes a group decision-making combination approach.
r eferences
Bartlett, M. S., Movellan, J. R., & Sejnowski, T. J. (2002). Face recognition by
independent component analysis. IEEE Transactions on Neural Networks, 13(6),
1450-1464.
Baudat, G., & Anouar, F. (2000). Generalized discriminant analysis using a kernel
approach. Neural Computation, 12(10), 2385-2404.
Belhumeur, P. N., Hespanha, J. P., & Kriengman, D. J. (1997). Eigenfaces vs. Fish-
erfaces: Recognition using class specifc linear projection. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 19(7), 711-720.
Belkin, M., & Niyogi, P. (2002). Laplacian eigenmaps and spectral techniques for
embedding and clustering. Advances in Neural Information Processing Systems,
14, 585-591.
Billings, S. A., & Lee, K. L. (2002). Nonlinear Fisher discriminant analysis using a
minimum squared error cost function and the orthogonal least squares algorithm.
Neural Networks, 15(2), 263-270.
Bolle, R. M., Connell, J. H., Pankanti, S., Ratha, N. K., & Senior, A. W. (2004).
Guide to biometrics. New York: Springer-Verlag.
Brunelli, R., & Falavigna, D. (1995). Person identifcation using multiple cues. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 12(10), 955966.
Overview 17
is prohibited.
Cawley, G. C., & Talbot, N. L. C. (2003). Effcient leave-one-out cross-validation of
kernel fsher discriminant classifers. Pattern Recognition, 36(11),2585-2592.
Chen, L. F., Liao, H. Y., Lin, J. C., Kao, M. D., & Yu, G. J. (2000). A new LDA-based
face recognition system which can solve the small sample size problem. Pattern
Recognition, 33(10), 1713-1726.
Draper, B. A., Baek, K., Bartlett, M. S., & Beveridge, J. R. (2003). Recognizing faces
with PCA and ICA. Computer Vision and Image Understandin,, 91(1/2), 115-137.
Duin, R. P. W., & Loog, M. (2004). Linear dimensionality reduction via a heterosce-
dastic extension of LDA: the Chernoff criterion. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 26(6), 732-739.
Fierrez-Aguilar, J., Ortega-Garcia, J., Gonzalez-Rodriguez, J., & Bigun, J. (2005).
Discriminative multimodal biometric authentication based on quality measure.
Pattern Recognition, 38(5), 777-779.
Gestel, T. V., Suykens, J. A. K., Lanckriet, G., Lambrechts, A., De Moor, B., & Van-
derwalle, J. (2002). Bayesian framework for least squares support vector machine
classifers, gaussian processs and kernel fsher discriminant analysis. Neural Com-
putation, 15(5), 1115-1148.
Gokberk, B., Salah, A. A., & Akarun, L. (2003). Rank-based decision fusion for
3D shape-based face recognition. Lecture Notes in Computer Science, 3546, 1019-
1028.
Grother, P., & Tabassi, E. (2004). NIST Biometric Scores Set - Release 1 (BSSR1).
Available at http://www.itl.nist.gov/iad/894.03/biometricscores.
Howland, P., & Park, H. (2004). Generalizing discriminant analysis using the gen-
eralized singular va Lu e decomposition. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 26(8), 995-1006.
Jain, A., Bolle, R., & Pankanti, S. (1999a). Biometrics: Personal Identifcation in
Networked Society. Boston Hardbound: Kluwer Academic.
Jain, A. K., Prabhakar, S., & Chen, S. (1999b). Combining multiple matchers for
a high security fngerprint verifcation system. Pattern Recognition Letters, 20,
1371-1379.
Jain, A. K., Ross, A., & Prabhakar, S. (2004). An introduction to biometric recogni-
tion. IEEE Trans. Circuits and Systems for Video Technology, 14(1), 4-20.
Jain, A. K., Nandakumar, K., & Ross, A. (2005). Score normalization in multimodel
biometric systems. Pattern Recognition, 38(5), 2270-2285.
is prohibited.
Jin, Z., Yang, J. Y., Hu, Z. S., & Lou, Z. (2001). Face Recognition based on uncor-
related discriminant transformation. Pattern Recognition, 34(7), 1405-1416.
Jing, X.Y., Zhang, D., & Yang, J. Y. (2003). Face recognition based on a group deci-
sion-making combination approach. Pattern Recognition, 36(7), 1675-1678.
Kim, H. C., Kim, D., Bang, S. Y., & Lee, S. Y. (2004). Face recognition using the
second-order mixture-of-eigenfaces method. Pattern Recognition, 37(2), 337-349.
Kirby, M., & Sirovich, L. (1990). Application of the KL procedure for the charac-
terization of human faces, IEEE Transactions on Pattern Analysis Machine Intel-
ligence, 12(1), 103-108.
Kong, A., Zhang, D., & Kamel, M. (2006). Palmprint identifcation using feature-
level fusion, Pattern Recognition, 39, 478-487.
Lathauwer, L., Moor, B., & Vndewalle, J. (2000). A multilinear singular value decom-
position. SIAM Journal on Matrix Analysis and Application, 21(4), 1233-1278.
Lawrence, N. D., & Schlkopf, B. (2001). Estimating a kernel Fisher discriminant
in the presence of label noise. Proceedings of 18th International Conference on
Machine Learning. San Francisco, CA, 306-313.
Liang, Z., & Shi, P. (2005). An analytical method for generalized low rank ap-
proximation matrices. Pattern Recognition, 38(11), 2213-2216.
Liang, Z., Zhang, D., & Shi, P. (2007). The theoretical analysis of GRLDA and its
applications. Pattern Recognition, 40(3), 1032-1041.
Liu, K., Cheng, Y. Q., Yang, J. Y., & Liu, X. (1992). An effcient algorithm for Foley-
Sammon optimal set of discriminant vectors by algebraic method. International
Journal of Pattern Recognition and Artifcial Intelligence, 6(5), 817-829.
Liu, K., Cheng, Y. Q., & Yang, J. Y. (1993). Algebraic feature extraction for im-
age recognition based on an optimal discriminant criterion. Pattern Recognition,
26(6), 903-911.
Liu, C., & Wechsler, H. (2000). Robust coding schemes for indexing and retrieval from
large face databases. IEEE Transactions on Image Processing, 9(1), 132-137.
Liu, C., & Wechsler, H. (2001). A shape- and texture-based enhanced Fisher classifer
for face recognition. IEEE Transactions on Image Processing, 10(4), 598-608.
Liu, C., & Wechsler, H. (2003). Independent component analysis of Gabor features
for face recognition. IEEE Transactions on Neural Networks, 14(4), 919-928.
Overview 19
is prohibited.
Liu, C. (2004). Gabor-based kernel PCA with fractional power polynomial models
for face recognition. IEEE Transactions on Pattern Analysis and Machine Intel-
ligence, 26(5), 572-581.
Loog, M., Duin, R. P. W., & Haeb-Umbach, R. (2001). Multiclass linear dimen-
sion reduction by weighted pairwise Fisher criteria. IEEE Transactions on Pattern
Lu, J. W., Plataniotis, K. N., & Venetsanopoulos, A. N. (2003a). Face recognition
using LDA-based algorithms. IEEE Transactions on Neural Networks, 14(1), 195-
200.
Lu, J., Plataniotis, K. N., & Venetsanopoulos, A. N. (2003b). Face recognition us-
ing kernel direct discriminant analysis algorithms. IEEE Transactions on Neural
Networks, 14(1), 117-126.
Lu, G., Zhang, D., & Wang, K. (2003). Palmprint Recognition using Eigenpalms
Features, Pattern Recognition Letters, 24(9-10), 1463-1467.
Miller, B. (1994). Vital signs of identity. IEEE Spectrum, 31(2), 22-30.
Mika, S., Rtsch, G., Weston, J., Schlkopf, B., & Mller, K. R. (1999a). Fisher dis-
criminant analysis with kernels. IEEE International Workshop on Neural Networks
for Signal Processing, Madison, Wisconsin, August 23-25 (pp. 41-48).
Mika, S., Rtsch, G., Schlkopf, B., Smola, A., Weston, J., & Mller, K. R. (1999b).
Invariant feature extraction and classifcation in kernel spaces. Advances in Neural
Information Processing Systems, 12, 526-532.
Mika, S., Rtsch, G., & Mller, K. R. (2001a). A mathematical programming ap-
proach to the Kernel Fisher algorithm. Advances in Neural Information Processing
Systems, 13, 591-597.
Mika, S., Smola, A. J., & Schlkopf, B. (2001b). An improved training algorithm
for kernel fsher discriminants. International Workshop on Artifcial Intelligence
and Statistics, San Francisco, CA, January 4-7, 98-104.
Mika, S., Rtsch, G., Weston, J., Schlkopf, B., Smola, A., & Mller, K. R. (2003).
Constructing descriptive and discriminative non-linear features: Rayleigh coeff-
cients in kernel feature spaces. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 25(5), 623-628.
Moghaddam, B. (2002). Principal manifolds and probabilistic subspaces for visual
recognition. IEEE Transactions on Pattern Analysis Machine Intelligence, 24(6),
780-788.
is prohibited.
Nandakumar, K., Chen, Y., Jain, A. K., & Dass, S. (2006). Quality-based score
level fusion in multibiometric systems. Proceedings of International Conference
on Pattern Recognition, Hong Kong, August 20-24 (pp. 473-476).
Pentland, A., Moghaddam, B., & Starner, T. (1994). View-based and mododular
eigenspaces for face recognition. IEEE Conference on Computer Vision and Pat-
tern Recognition, Seattle, June 21-23 (pp. 84-91).
Pentland, A. (2000). Looking at people: sensing for ubiquitous and wearable comput-
ing. IEEE Transactions on Pattern Analysis Machine Intelligence, 22(1), 107-119.
Petridis, S., & Perantonis, S. T. (2004). On the relation between discriminant
analysis and mutual information for supervised linear feature extraction. Pattern
Recognition, 37(5), 857-874.
Poh, N., & Bengio, S. (2006). Database, protocol and tools for evaluating score-
level fusion algorithms in biometric authentication. Pattern Recognition, 39(2),
223-233.
Ross, A., & Govindarajan, R. (2005). Feature level fusion using hand and face
biometrics. Proceedings of SPIE Conference on Biometric Technology for Human
Identifcation, Orlando, USA, March 28 (pp. 196-204).
Roth, V., & Steinhage, V. (2000). Nonlinear discriminant analysis using kernel
functions. Advances in Neural Information Processing Systems, 12, 568-574.
Roweis, S. T., & Saul, L. K. (2000). Nonlinear dimensionality reduction by locally
linear embedding. Science, 290, 2323-2326.
Schlkopf, B., Smola, A., & Muller, K. R. (1998). Nonlinear component analysis as
a kernel eigenvalue problem. Neural Computation, 10(5), 1299-1319.
Sirovich, L., & Kirby, M. (1987). Low-dimensional procedure for characterization
of human faces. Journal of Optical Society of America A, 4, 519-524.
Swets, D. L., & Weng, J. (1996). Using discriminant eigenfeatures for image retrieval.
IEEE Transactions on Pattern Analysis and Machine Intelligence, 18(8), 831-836.
Tao, D., Li, X., Hu, W., Maybank, S. J., & Wu, X. (2005). Supervised tensor learn-
ing. IEEE International Conference on Data Mining, Houston, Texas, November
27-30 (pp. 450-457).
Tenenbaum, J. B., de Silva, V., & Langford, J. C. (2000). A global geometric frame-
work for nonlinear dimensionality reduction. Science, 290, 2319-2323.
Tian, Q., Barbero, M., Gu, Z. H., & Lee, S. H. (1986). Image classifcation by the
Foley-Sammon transform. Optical Engineering, 25(7) 834-839.
Overview 21
is prohibited.
Turk, M., & Pentland, A. (1991a). Eigenfaces for recognition. Journal of Cognitive
Neuroscience, 3(1), 71-86.
Turk, M., & Pentland, A. (1991b). Face recognition using Eigenfaces. IEEE Confer-
ence on Computer Vision and Pattern Recognition, Maui, Hawaii, June 3-6 (pp.
586-591).
Ulery, B., Hicklin, A. R., Watson, C., Fellner, W., & Hallinan, P. (2006). Studies of
biometric fusion. NIST Technical Report IR7346.
Vasilescu, M. A. O., & Terzopoulos, D. (2003). Multilinear subspace analysis of
image ensembles. IEEE Conference on Computer Vision and Pattern Recognition,
Madison, Wisconsin, June 16-22 (pp. 93-99).
Wayman, J. (2005). Biometric systems: Technology, design and performance evalu-
ation. London: Springer.
Wu, X., Zhang, D., & Wang, K. (2003). Fisherpalms based Palmprint recognition.
Pattern Recognition Letters, 24(15), 2829-2838.
Xu, J., Zhang, X., & Li, Y. (2001). Kernel MSE algorithm: a unifed framework for
KFD, LS-SVM, and KRR. Proceedings of the International Joint Conference on
Neural Networks, Washington, D.C., July 15-19 (pp. 1486-1491).
Xu, Y., Yang, J. Y., & Yang, J. (2004). A reformative kernel Fisher discriminant
analysis. Pattern Recognition, 37(6), 1299-1302.
Yan, S., Xu, D., Yang, Q., Zhang, L., Tang, X., & Zhang, H. J. (2007). Multilinear
discriminant analysis for face recognition. IEEE Transactions on Image Process-
ing, 16(1), 212-220.
Yang, M. H. (2002). Kernel Eigenfaces vs. kernel Fisherfaces: Face recognition
using kernel methods. Proceedings of the Fifth IEEE International Conference on
Automatic Face and Gesture Recognition (RGR02), Washington D. C., May 21
(pp. 215-220).
Yang, J., & Yang, J. Y. (2002). Generalized K-L transform based combined feature
extraction. Pattern Recognition, 35(1), 295-297.
Yang, J., & Yang, J. Y. (2003). Why can LDA be performed in PCA transformed
space? Pattern Recognition, 36(2), 563-566.
Yang, J., Yang, J. Y., Zhang, D., & Lu, J. F. (2003a). Feature fusion: parallel strategy
vs. serial strategy. Pattern Recognition, 36(6), 1369-1381.
is prohibited.
Yang, J., Yang, J. Y., Frangi, A. F. , & Zhang, D. (2003b). Uncorrelated projection dis-
criminant analysis and its application to face image feature extraction. International
Yang, J.,

Jin, Z., Yang, J. Y., Zhang, D., & Frangi, A. F. (2004a). Essence of Kernel
Fisher Discriminant: KPCA plus LDA. Pattern Recognition, 37(10), 2097-2100.
Yang, J.,

Frangi, A. F., &

Yang, J. Y. (2004b). A new kernel Fisher discriminant
algorithm with application to face recognition, Neurocomputing, 56, 415-421.
Yang, J., Zhang, D., Frangi, A. F., & Yang, J. Y.

(2004c). Two-dimensional PCA: A
new approach to face representation and recognition. IEEE Transactions on Pattern
Yang, J., Zhang, D., Yang, J. Y., Zhong, J., & Frangi, A. F. (2005). KPCA plus
LDA: A complete kernel Fisher discriminant framework for feature extraction and
recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence,
27(2), 230-244.
Ye, J. (2004). Generalized low rank approximations of matrices. The Twenty-First
International Conference on Machine Learning, Bonn, Germany, August 7-11.
Ye, J., Janardan, R., Park, C. H., & Park, H. (2004). An optimization criterion for
generalized discriminant analysis on undersampled problems. IEEE Transactions
on Pattern Analysis and Machine Intelligence, 26(8), 982-994.
Yu, H., & Yang, J. (2001). A direct LDA algorithm for high-dimensional datawith
application to face recognition. Pattern Recognition, 34(10), 2067-2070.
Yuen, P. C., & Lai, J. H. (2002). Face representation using independent component
analysis. Pattern Recognition, 35(6), 1247-1257.
Zhang, D. (2000). Automated biometrics: tecnologies and systems. Boston Hard-
bound: Kluwer Academic.
Zhang, D. (2002). Biometrics solutions for authentication in an e-world. Boston
Hardbound: Kluwer Academic.
Zhang, D. (2004). Palmprint authentication. Boston Hardbound: Kluwer Academic
Publishers.
Zhang, D., Peng, H., Zhou, J., & Pal, S. K. (2002). A novel face recognition system
using hybrid neural and dual eigenfaces methods. IEEE Transactions on System,
Man, and Cybernetics, Part A, 32(6), 787-793.
Zhang, D., Kong, W. K., You, J., & Wong, M. (2003). On-line palmprint identifcation.
IEEE Transactions on Pattern Analysis Machine Intelligence, 25(9), 1041-1050.
Overview 23
is prohibited.
Zhao, L., & Yang, Y. (1999). Theoretical analysis of illumination in PCA-based
vision systems. Pattern Recognition, 32(4), 547-564.
Zhao, W., Krishnaswamy, A., Chellappa, R., Swets, D., & Weng, J. (1998). Dis-
criminant analysis of principal components for face recognition. In H. Wechsler,
P.J. Phillips, V. Bruce, F.F. Soulie , & T.S. Huang (Eds.), Face Recognition: From
Theory to Applications 73-85, Berlin: Springer-Verlag.
Zhao, W., Chellappa, R., & Rosenfeld, A. (2003). Face recognition: a literature
survey. ACM Computing Survey, 36(4), 399-458.
Zuo, W., Wang, K., & Zhang, D. (2005). Bi-directional PCA with assembled matrix
distance metric. International Conference on Image Processing Genova, Italy,
September 11-14 (pp. 958-961).
Zuo, W., Wang, K., Zhang, D., & Zhang, H. (2007). Combination of two novel LDA-
based methods for face recognition. Neurocomputing, 70(4-6), 735-742.

Section I
Biometric Discriminant Analysis
Discriminant Analysis for Biometric Recognition 25
is prohibited.
Chapter II
Discriminant Analysis for
Biometric Recognition
Abstr Act
This chapter is a brief introduction to biometric discriminant analysis technologies
Section I of the book. Section 2.1 describes two kinds of linear discriminant analy-
sis (LDA) approaches: classifcation-oriented LDA and feature extraction-oriented
LDA. Section 2.2 discusses LDA for solving the small sample size (SSS) pattern
recognition problems. Section 2.3 shows the organization of Section I.
2.1 lI ne Ar dI scr ImInAnt AnAl ys Is
Linear discriminant analysis (LDA) method has been widely studied in and suc-
cessfully applied to biometric recognition such as face, fngerprint, and palm print
identifcation or verifcation.
The essence of LDA is to construct a linear discriminant criterion which can
be used to build a binary classifer or a feature extractor. To differentiate LDA for
binary classifcation from LDA for feature extraction, hereafter we name the former
as classifcation-oriented LDA, and the later feature extraction-oriented LDA.
2.1.1 Classi.cation-Oriented Linear Discriminant Analysis
Linear discriminant analysis was initially developed for binary classifcation in
the seminal work of LDA (Fisher, 1936). Among various discriminant criteria, one
is prohibited.
of the most famous is Fisher discriminant criterion (FDC) for binary linear dis-
criminant analysis. FDC tries to seek an optimal projection direction such that the
between-class variance is maximized while the within-class variance is minimized
if samples from two distinct classes are projected along this projection direction.
Besides FDC, there exist other linear discriminantcriteria for binary classifcation.
Among them, perceptron and minimum squared-error (MSE) (Duda, Hart, & Stork,
2001) criteria are two well-known examples.
The mathematical form of each linear discriminant criterion can be character-
ized as an optimization model which is used to calculate the weight and sometimes
the bias of a binary classifer. Similarly, the mathematical form of linear support
vector machine (LSVM) (Burges, 1998) is also an optimization model to calculate
the weight and bias for a binary classifer. Thus, LSVM can be viewed as a kind of
classifcation-oriented LDA method.
Classifcation-oriented LDA methods are in fact binary linear classifers and
can not directly be applied to multiple pattern classifcation tasks. They should be
used in combination with one of the implementation strategies described in Section
3.1.3 if they are applied to these tasks.
2.1.2 Feature Extraction-Oriented Linear Discriminant Analysis
Based on Fishers work, Wilks (1962) extended the concept of optimal projection
direction to a set of discriminant vectors by extending Fisher discriminant criterion
to multiple Fisher discriminant criterion. While the former is used to calculate the
weight of the Fisher classifer, the latter is used to compute a set of Fisher discriminant
vectors. By using these discriminant vectors as a transformation matrix, Wilks suc-
cessfully reduced a complicated classifcation problem in a high-dimensional input
space into a simple one in a low-dimensional feature space. The procedure which
compresses the data from the input space into a feature space by utilizing a trans-
formation matrix is called linear feature extraction. The feature extraction method
using Fisher discriminant vectors is called Fisher linear discriminant (FLD).
In general, Fisher discriminant vectors are unnecessarily orthogonal to each
other. Many researchers believed that the discriminant capability of FLD could be
enhanced by removal of linear dependence among Fisher discriminant vectors. Based
on this intuition, Foley-Sammon discriminant (FSD) a feature extraction method
which used a set of orthogonal discriminant vectors was subsequently developed
(Sammon, 1970; Foley & Sammon, 1975; Duchene & Leclercq, 1988).
is prohibited.
2.2 ld A for sol vIng the smAll sAmPle sI ze
Problems
2.2.1 t he small sample size Pattern r ecognition Problems
Since FDC, MSE, FLD, and FSD all involve the computation of the inverse of one
or several scatter matrices of sample data, it is a precondition that these matrices
should be nonsingular. In the small sample size (SSS) pattern recognition problems
such as appearance-based face recognition, the ratio of dimensionality of input space
to the number of samples is so large that the matrices involved are all singular. As
a result, standard LDA methods cannot directly be applied to these SSS problems.
Due to the prospective applications to biometric identifcation and computer vision,
LDA for solving the SSS problems becomes one of the hottest research topics in
pattern recognition.
2.2.2 Studies on Classifcation-Oriented LDA for SSS Problems
In comparison with rich literature in feature extraction-oriented LDA for SSS
problems, studies on pattern classifcation aspect of LDA for SSS problems are
quite few. To the best of our knowledge, except for large margin linear projection
(LMLP) (Song, Yang, & Liu, 2004a), minimum norm minimum squared-error
(MNMSE) (Song, Yang & Liu, 2004b), and maximum scatter difference (MSD)
(Song, Cheng, Yang & Liu, 2004; Song, Zhang, Chen & Wang, 2007) there is almost
no endeavor in this direction.
2.2.3 Studies on Feature Extraction-Oriented LDA for SSS
Problems
Linear discriminant feature extraction methods for solving the SSS problems
generally fall into two categories. One is computation-oriented; the other is sub-
space-based. Some researchers think that the trouble that standard LDA methods
encounter in the SSS problems is only computational and can be avoided by using
approximated calculation. Typical computation-oriented methods include the pseudo-
inverse method (Tian, Barbero, Gu & Lee, 1986), the singular value disturbance
method (Hong & Yang, 1991), and the regularization method (Zhao, Chellappa, &
Phillip, 1999).
By treating the standard LDA method as an ill-posed optimization model with
respect to SSS problems, many researchers focus their attentions on the modifca-
tion of the optimization model corresponding to the LDA method, especially on the
feasible region of the optimization model. Representative subspace-based methods
is prohibited.
include Fisherface (Belhumeur, Hespanha & Kriengman, 1997), nullspace LDA
(N-LDA) (Chen, Liao, Ko, Lin & Yu, 2000), direct LDA (D-LDA) (Yu & Yang,
2001), and complete LDA (C-LDA) (Yang & Yang, 2003). Belhumeur Hespanha
and Kriengman (1997) tried to seek discriminant vectors in a subspace of the
range of the total scatter matrix. Chen, Liao, Ko, Lin and Yu (2000) tried to seek
discriminant vectors in the null space of the within-class scatter matrix. Yu and
Yang (2001) tried to seek discriminant vectors in the range of the between-class
scatter matrix. Yang and Yang (2003) tried to seek discriminant vectors both in the
range of the between-class scatter matrix and in the null space of the within-class
scatter matrix.
2.3 org AnIzAt Ion of sect Ion I
The frst section of the book presents some of our recent studies on LDA for solv-
ing the SSS problems and its applications for biometric recognitions. It consists of
fve chapters. Chapter II provides a brief introduction. Chapter III describes our
researches on the pattern classifcation aspect of LDA for SSS problems. Chapter
IV-VI present our studies on the feature extraction aspect of LDA for SSS problems.
More specifcally, Chapter IV focuses on orthogonal discriminant analysis methods.
Chapter V discusses parameterized discriminant analysis methods, and Chapter VI
describes two novel facial feature extraction methods, multiple maximum scatter
difference and discriminant based on coeffcients of variances.
r eferences
Belhumeur, P. N., Hespanha, J. P., & Kriengman, D. J. (1997). Eigenfaces vs. fsher-
faces: recognition using class specifc linear projection. IEEE Trans. Pattern Anal.
Machine Intell., 19(7), 711-720.
Burges, C. J. C. (1998). A tutorial on support vector machines for pattern recogni-
tion. Data Mining and Knowledge Discovery, 2(2), 121167.
Chen, L., Liao, H., Ko, M., Lin, J., & Yu, G. (2000). A new lda-based face recogni-
tion system which can solve the small sample size problem. Pattern Recognition,
33(10), 1713-1726.
Duchene, J., & Leclercq, S. (1988). An optimal transformation for discriminant and
principal component analysis. IEEE Trans. Pattern Anal. Machine Intell., 10(6),
978-983.
is prohibited.
Duda, R. O., Hart, P. E., & Stork, D. G. (2001). Pattern Classifcation. New York:
John Wiley & Sons.
Fisher, R. (1936). The use of multiple measurements in taxonomic problems. Annals
of Eugenics, 7(2), 178-188.
Foley, D. H., & Sammon, J. W. (1975). An optimal set of discriminant vectors. IEEE
Trans. Computer, 24(3), 281-289.
Hong, Z. Q., & Yang, J. Y. (1991). Optimal discriminant plane for a small number
of samples and design method of classifer on the plane. Pattern Recognition, 24(4),
317-324.
Sammon, J. W. (1970). An Optimal discriminant plane. IEEE Trans. Computer,
C-19(9), 826-829.
Song, F. X., Cheng, K., Yang, J. Y., & Liu, S. H. (2004). Maximum scatter differ-
ence, large margin linear projection and support vector machines. Acta Automatica
Sinica (in Chinese), 30(6), 890-896.
Song, F. X., Yang, J. Y., & Liu, S. H. (2004a). Large margin linear projection and
face recognition. Pattern Recognition, 37(9), 1953-1955.
Song, F. X., Yang, J. Y., & Liu, S. H. (2004b). Pattern recognition based on the
minimum norm minimum squared-error classifer. Proceedings of the Eighth
International Conference on Control, Automation, Robotics and Vision (ICARCV
2004), (pp. 1114-1117).
Song, F. X., Zhang, D., Chen, Q. L., & Wang, J. Z. (2007). Face recognition based
on a novel linear discriminant criterion. Pattern Analysis & Applications, 10(3),
165-174.
Foley-Sammon transform. Optical Engineering, 25(7), 834-840.
Wilks, S. S. (1962). Mathematical Statistics, New York: Wiley.
Yang, J., & Yang, J. Y. (2003). Why can lda be performed in pca transformed space?
Yu, H., & Yang, J. (2001). A direct LDA algorithm for high-dimensional datawith
Zhao, W., Chellappa, R., & Phillip, P. J. (1999). Subspace linear discriminant
analysis for face recognition (CAR-TR-914). University of Maryland, Center for
Automatic Research.
is prohibited.
Chapter III
Discriminant Criteria for
Pattern Classifcation
Abstr Act
As mentioned in Chapter II, there are two kinds of LDA approaches: classifca-
tion-oriented LDA and feature extraction-oriented LDA. In most chapters of this
session of the book, we focus our attention on the feature extraction aspect of LDA
for SSS problems. On the other hand,, with this chapter we present our studies on
the pattern classifcation aspect of LDA for SSS problems. In this chapter, we pres-
ent three novel classifcation-oriented linear discriminant criteria. The frst one is
large margin linear projection (LMLP) which makes full use of the characteristic
of the SSS problems. The second one is the minimum norm minimum squared-er-
ror criterion which is a modifcation of the minimum squared-error discriminant
criterion. The third one is the maximum scatter difference which is a modifcation
of the Fisher discriminant criterion.
3.1 Introduct Ion
3.1.1 Linear Discriminant Function and Linear Classifer
Let
1
,...,
d
N
R x x be a set of training samples from two classes
1
and
2
with N
i

samples from
i
, and let
1
,..., {1, 1}
N
y y - be their corresponding class labels. Here
y
j
= 1 means that
j
x belongs to
1
whereas y
j
= 1 means that
j
x belongs to
2
. A
Discriminant Criteria for Pattern Classi.cation 31
is prohibited.
linear discriminant function is a linear combination of the components of a feature
vector
d
R x which can be written as:
0
( )
T
g w = + x w x (3.1)
where the vector
d
R w and the scalar w
0
are called weight and bias respectively.
The hyperplane
0
0
T
w + = w x is a decision surface which is used to separate samples
with positive class labels from samples with negative ones.
A linear discriminant criterion is an optimization model which is used to seek the
weight for a linear discriminant function. The chief goal of classifcation-oriented
LDA is to set up an appropriated linear discriminant criterion and to calculate the
optimal projection direction, i.e. the weight. Here optimal means that after samples
are projected onto the weight, the resultant projections of samples from two distinct
classes
1
and
2
are fully separated.
Once the weight
*
w has been derived from a certain linear discriminant criterion,
the corresponding bias w
0
can be computed using:
*
0
T
w = -w m (3.2)
or
* 1 2
0
2
T
w
+
| |
= -
|
\ .
m m
w (3.3)
where m and
i
m are respectively the mean training sample and the mean of training
samples from the class
i
. They are defned as
1 2
1
N

=

x
m x
, (3.4)
and
1
i
i
i
N

=

x
m x. (3.5)
For simplicity, we calculate the bias using the Eq. (3.2) throughout this chap-
ter.
Let
1
* *
1 1
1
1
T T
m
N

= =
x
w x w m
is prohibited.
denote the mean of the projected training samples from the class
1
. Thus, the binary
linear classifer based on the weight
*
w and the bias w
0
is defned as follow:
*
0 1 0
( ) ( ) ( )
T
f sign w sign m w = + + x w x
, (3.6)
which assigns a class label
*
0 1 0
( ) ( )
T
sign w sign m w + + w x to an unknown sample
x. Here, sign is the sign function. That is, once the weight in a linear discriminant
function has been worked out the corresponding binary linear classifer is fxed.
3.1.2 l inear support vector machine
Linear support vector machine (LSVM) (Burges, 1998) is one of the top perform-
ing classifers in the feld of pattern recognition. It has been successfully applied
to text mining, machine learning, and computer vision. LSVM is one of the most
important statistical learning methods which are based on the structural risk mini-
mization principle (Vapnik, 1995). It is a binary classifer in nature and tries to
seek an optimal hyperplane which maximizes the margin between samples from
two distinct classes.
When training samples are linearly separable, the weight w and the bias b of the
optimal decision superplane 0
T
b - = w x are calculated by the following model:
2 1
min
2
( ) 1
T
i i
y b -
w
w x . (3.7)
If training samples are linearly inseparable, the weight w and the bias b of the
optimal separated superplane can be derived from the following model:
2
1
1
min
2
( ) 1
0, 1,...,
N
i
i
T
i i i
i
C
y b
i N
=
+
- -
=
w
w x
(3.8)
Here
i
,
1,..., i N =
are slack variables and the parameter C controls the tradeoff
between large margin and small training error. C has only positive values.
Discriminant Criteria for Pattern Classifcation 33
is prohibited.
3.1.3 Using Binary Classifers for Multiclass Classifcation
It is well known that binary classifers cannot be directly applied to multiclass clas-
sifcation problems. To use binary classifers for multiclass classifcation we have
to divide a multiclass problem into a series of binary classifcation problems using
one of the three implementation strategies: one-vs-rest, one-vs-one, and directed-
acyclic-graph (Hsu & Lin, 2002).
In the one-vs-rest strategy, an l-class problem is divided into l binary problems.
In the ith problem a binary classifer is trained with all samples from the ith class
with positive class labels, and all other samples with negative class labels. Thus
we obtain l linear discriminant functions
0
( )
T i
i i
g w = + x w x ,
1,..., i l =
. An unknown
sample x is assigned to the jth class if
1
( ) max ( )
j i
i l
g g

= x x .
In the one-vs-one strategy, an l-class problem is divided into ( 1) 2 l l - binary
problems: P(1, 2), , P (i, j), , P (l-1, l). In the problem P (i, j), a binary classi-
fer is trained with all samples from the ith class with positive class labels, and all
samples from the jth class with negative class labels. This classifer assigns each
unknown sample x to the ith class or the jth class according to the output value of
the learned linear discriminant function. Thus, each unknown sample x holds a set
of ( 1) 2 l l - class labels. The fnal class label of the sample x is j if j appears most
frequently in the set of class label of x.
The training phase of the directed-acyclic-graph strategy is the same as the
one-vs-one strategy by training ( 1) 2 l l - binary classifers. However, in the test-
ing phase, it uses a rooted binary directed acyclic graph(see Figure 3.1) which has
( 1) 2 l l - internal nodes and l leaves. Each node is a binary classifer of the ith and
Figure 3.1. Rooted binary directed acyclic graph
(1, l)
(1, l-1) (2, l)
(1, l-2) (2, l-1) (3, l)

1 2 3
l-1 l

is prohibited.
the jth classes. Given an unknown sample x, starting at the root node, the binary
linear discriminant function is evaluated. Then it moves to either left or right de-
pending on the output value. Therefore, we go through a path before reaching a
leaf node which indicates the predicted class. Since numerous studies have shown
that the one-vs-one implementation strategy is most effective, this implementation
strategy is used in this book.
3.2 lA rge mArg In lI ne Ar Project Ion cl Ass If Ier
3.2.1 f isher discriminant criterion
The Fisher discriminant criterion (FDC) (Duda, Hart, & Stork, 2001) tries to seek
an optimal projection direction such that along that direction the within-class vari-
ance of projected training samples is minimized while the between-class variance
of projected samples is maximized. The weight derived from FDC is called Fisher
weight, and is denoted as w
F
.
Specifcally, the FDC maximizes the generalized Rayleigh quotient (Duda, Hart
and Stork, 2001):
( )
T
b
f
T
w
S
J
S
=
w w
w
w w
, (3.9)
where S
b
and S
w
are respectively the between- and within-class scatter matrices for
binary classifcation problems, which are defned as:
1 2 1 2
( )( )
T
b
S = - - m m m m
, (3.10)
2
1
( )( )
i
T
w i i
i
S
=
= - -
x
x m x m
, (3.11)
The Fisher weight can be computed in closed-form using the formula:
1
2
)
F w
S
-
= -
1
w (m m
, (3.12)
if the within-class scatter matrix S
w
is nonsingular.
In the SSS problems such as appearance-based face recognition, the matrix S
w

is always singular. Thus, how to deal with the singularity of the within-class scatter
matrix is one of the important problems in the feld of LDA (Belhumeur, Hespanha,
& Kriengman, 1997).
is prohibited.
3.2.2 Model and Algorithm
If the within-class scatter matrix S
w
is singular, there is at least one nonzero vector
w such that 0
T
w
S = w w . In this case, max ( )
f
J w becomes an ill-posed optimiza-
tion problem. It is natural to transform the optimization model of FDC into the
following one:
0
max
T
w
T
b
T
S
S
= w w
w w
w w
. (3.13)
The discriminant criterion corresponding to this optimization model is called
large margin linear projection (LMLP) (Song, Yang, & Liu, 2004). The weight
derived from LMLP is called LMLP weight, and is denoted as w
L
.
The LMLP weight w
L
can be computed in two distinct ways. We frst present
the frst approach.
Let
1
,...,
n
v v be an orthonormal basis for the nullspace of the within-class scat-
ter matrix S
w
, N(S
w
). According to Theorem 3 in Liu, Cheng and Yang (1992), the
constraint 0
T
w
S = w w is equivalent to ( )
w
N S w . As a result, the optimization
model (3.13) can be rewritten as
max
T T
b
T
V S V u u
u u
. (3.14)
Here
1
[ ,..., ]
n
V = v v . Using the Lagrangian multiplier method (Bian & Zhang,
2000), the LMLP weight can be calculated using the formula
1 L
V = w u
. (3.15)
Here
1
u
is the eigenvector of the matrix
T
b
V S V
corresponding to the largest
eigenvalue.
Since solving an orthonormal basis for the nullspace of a large scale matrix is
extremely time-expensive, the LMLP weight is seldom calculated using the formula
(3.15).
Now we discuss the second approach for computing the LMLP weight.
From the constraint 0
T
w
S = w w in the optimization model (3.13) and the defni-
tion of S
w
in (3.11), it is easy to conclude that there are at most two distinct projected
points m
1
and m
2
, and each of which corresponds to one class of samples, after all
is prohibited.
the samples are projected onto w. Since
2
1 2
2
( )
T
b
T
S m m -
=
w w
w w
w
,
thus maximizing
T
b
T
S w w
w w

is maximizing the margin or the distance between the projected samples from the
two classes.
Let
1
1
N d
Q R

and
2
2
N d
Q R

denote sample matrices for classes
1
and
2
re-
spectively. The model (3.13) can be reformulated as:
2
1 2
max( ) m m -
, subject to,
1
2
1 1
2 2
1
N
N
Q m
Q m
=
w 1
w 1
w
(3.16)
where
n
1 is a n-dimensional column vector whose elements are all constant 1.
Without losing generality, let
1 2
m m m = - = , and substitute w with / m w , then
the optimal model described by (3.16) can be further converted to:
2 1
min
2
w
, subject to
1
2
1
2
N
N
Q
Q
=
= -
w 1
w 1
. (3.17)
Thus, we can rewrite the optimization model again as:
2
1
1
min
2
T
i i
y = w x
w
. (3.18)
It is a quadratic programming problem which can be solved using various ready
algorithms effciently.
In fact, LMLP is a special type of LSVM which assumes that training samples
are linearly separable. Thus LMLP inherits the excellent properties of LSVM.
The feasible region of the quadratic programming model (3.18) is nonempty if
and only if training samples satisfy the following two conditions:
Consistentcondition If
i j
= x x , then
i j
y y = ;
Separablecondition If Q denotes the feature by sample matrix, then ( ) rank Q d .
is prohibited.
Table 3.1 Comparison of the best recognition rates (%) of the three classical LDA
techniques with that of LMLP
The consistent condition is a precondition for any pattern recognition problem.
If there do exist two samples which have the same feature vector but different class
labels, they should be removed from the sample data. In SSS problems the separable
condition is always true.
3.2.3 experimental evaluation
The performance of LMLP is tested on the ORL face image database. ORL database
is provided by Olivetti research lab (ORL, http://www.cam-or.co.uk/Facedatabase.
html). This set of data consists of 40 distinct persons, with each containing 10 dif-
ferent images with variation in pose, illumination, facial expression (open/closed
eyes, smiling/not smiling) and facial details (glasses / no glasses). All the images
are taken against a dark homogeneous background with the subjects in an upright,
frontal position with a tolerance for some tilting and rotation of the face of up to 20
degrees. Moreover, there is also some variation in the scale of up to about 10 percent.
All images were taken in grayscale with a 256-level and normalized to a resolution
of 11292 pixels. Five sample images of one individual from the ORL database are
shown in Figure 3.2. In this experiment, the frst fve images of each individual
are used for training and the remaining fve for testing. Thus the total amount of
the training samples and the testing samples are both 200. Table 3.1 shows the best
recognition rates obtained by using the nearest neighbor classifer in combination
with each of the three important facial feature extraction methods: Fisherface (Bel-
humeur, Hespanha, & Kriengman, 1997), Eigenfaces (Turk & Pentland, 1991), and
uncorrelated linear discriminant analysis (U-LDA) (Jin, Yang, Hu, & Lou, 2001)
as well as the recognition rate obtained by using LMLP classifer.
To further compare the performance of LMLP with that of LSVM, the best
recognition rates, the worst recognition rates, the mean recognition rates, and the

Figure 3.2. Sample images from the ORL face image database
Fisherface Eigenface Uncorrelated discriminant LMLP
88.5 93.5 88.0 94.5
is prohibited.
standard deviation of the two binary linear classifers based on 252 runs are shown
in Table 3.2. The fve different resolutions considered in the experiment are 11292,
5646, 2823, 1412 and 76 obtained as in Jin, Yang, Hu and Lou (2001). The
results listed in Table 3.2 are derived from
5
10
252 ( ) C = distinct runs for each reso-
lution. In each of the tests the ith, jth, kth, lth, and mth images of each individual
are used for training and the remaining fve for testing (1 10) i j k l m < < < < .
The total time for training and testing consuming by LMLP and LSVM for various
resolutions is shown in Table 3.3. From these experimental results we can conclude
that LMLP is a promising binary classifer for face recognition.
3.3 mInImum norm mInImum squ Ared- error
cl Ass If Ier
3.3.1 Minimum Squared-Error Criterion
Let
1
,...,
d
N
R x x be a set of d-dimensional training samples from two classes
1

and
2
with N
i
samples from
i
, and
1
,..., {1, 1}
N
y y - to be their class labels. Let
[ 1]
T T
i i i
y = y x
be the normalized augmented vector of
i
x , Y be the N-by-d matrix (
Table 3.2. Comparison of the performance of LMLP with that of LSVM
Table 3.3. The total time (sec) for training and classifcation consuming by LMLP
and LSVM
11292 5646 2823 1412 76
LMLP 85.9 25.5 10.1 7.0 6.0
LSVM 89.5 28.2 12.9 10.1 10.3
resolution 11292 5646 2823 1412 76
C
l
a
s
s
i
f
e
r
LSVM
max 99.0 99.0 99.0 99.0 99.0
min 90.5 90.5 91.0 90.5 89.5
mean 95.9 96.0 96.2 95.9 95.8
std 1.64 1.58 1.55 1.57 1.82
LMLP
max 99.5 99.0 99.5 99.0 99.0
min 91.5 91.5 91.5 91.0 89.5
mean 95.6 95.8 96.2 96.1 95.7
std 1.54 1.47 1.41 1.41 1.68
is prohibited.
1 d d = + ) whose ith row is the transpose of the vector

i
y , and b be a vector composed
of positive numbers
1
( ,..., )
T
N
b b = b . The minimum squared-error (MSE) criterion
tries to calculate the weight and the bias of a linear discriminant function by seek-
ing a solution to a set of linear equations
Y = a b. (3.19)
If Y is nonsingular, there is a unique solution
1
Y
-
= a b. When Y is singular or
rectangular, the inverse matrix of Y does not exist. As an alternative choice, MSE tries
to seek a vector which minimizes the following discriminant criterion function
2
2
1
( ) ( )
N
T
s i i
i
J Y b
=
= - = -
a a b a y
. (3.20)
Once the optimal vector
a
has been worked out, the weight w and the bias
0
w
are fxed by
0
[ ]
T T
w = w a. (3.21)
If the
d-by-
d matrix
T
Y Y is nonsingular, we can solve for a uniquely as
1
( )
T T
Y Y Y
-
= a b. (3.22)
But if the matrix
T
Y Y is singular, as is often the case in the SSS problems, there
are infnitely many solutions a such that ( ) 0
s
J = a . In these cases conventional MSE
solutions can not assure the optimal decision surface.
3.3.2 Minimum Norm Minimum Squared-Error Weight
In terms of matrix algebra, a vector a which minimizes the discriminant criterion
function ( )
s
J a is a MSE solution for the matrix equation (3.19). No matter whether
the vector a is over-determined or underdetermined by Eq. (3.19), a MSE solution
for it always exists.
While there might be infnite MSE solutions, the minimum norm minimum
squared-error solution is unique
Y
+
= a b, (3.23)
where Y
+
is the Moore-Penrose inverse of the matrix Y.
Y
+
can be computed in two ways (Cheng, 1989):
is prohibited.
1.
1 T
Y VS U
+ -
= , if the singular value decomposition of Y is USV
T
;
2.
1
( )
T T
Y V U YV U
+ -
= , if the rank decomposition of Y is UV
T
.
Let
N
1 be the column vector whose elements are all ones. The minimum norm
minimum squared-error (MNMSE) weight is fxed by
0
[ ]
T T
N
w Y
+
= w 1
. (3.24)
The physical explanation of the MNMSE weight is presented asfollowing.
When there are fewer equality constraints than unknown variables and con-
straints are consistent in Eq. (3.19), a MSE solution degenerates into a solution of
the underdetermined linear system.
Let w be the weight corresponding to a solution a for
N
Y = a 1 . Thus, along the
direction w all samples from the class
1
are projected to the point 1 and all samples
from the class
2
are projected to the point -1, i.e.
1
1
T
T
+
-
=
= -
w x
w x , (3.25)
where
+
x and
-
x stand for a sample from
1
and
2
respectively.
It is easy to understand that the margin between the two projected points is 2/ w .
Therefore while all MSE solutions separate the positive samples from the negative
ones correctly, the MNMSE weight separates them with the largest margin.
To properly evaluate the performance of MNMSE a series of experiments are car-
ried out in this subsection. The frst experiment is conducted on the Yale face image
database. The Yale face image database contains 165 images of 15 individuals (each
person has 11 different images) under various facial expressions and lighting con-
ditions. All images are grayscale and normalized to a resolution of 10080 pixels.
Sample images of one person from the Yale database are shown in Figure 3.3.
Besides LSVM and LMLP, four facial feature extractor: Eigenface, independent
component analysis (ICA) (Bartlett, Movellan, & Sejnowski, 2002), two-dimensional
principal component analysis (2DPCA) (Yang, Zhang, Frangi, & Yang, 2004), and
kernel Eigenface (Yang, 2002) are also evaluated in the experiment. The leave-one-
out evaluation strategy is adopted, that is, each sample is used as a testing sample
while the others are used as training samples. Again, the nearest neighbor classifer
is used in combination with feature extraction methods. Experimental results are
is prohibited.
Figure 3.3. Sample images of one person from Yale face database
Table 3.4. Comparison of the recognition rates of MNMSE with those of other face
recognition techniques using Yale database
Method PCA ICA KPCA 2DPCA LSVM LMLP MNMSE
Accuracy 71.52%
*
71.52%
*
72.73%
*
84.24%
90.91% 92.73% 92.73%

listed in Table 3.4. Here PCA stands for Eigenface and KPCA stands for Kernel
Eigenface. The results with symbol * are from (Yang et al., 2004). The result
with symbol is from (Bartlett, Movellan, & Sejnowski, 2002). It is showed that
MNMSE is superior to all of face recognition techniques except for LMLP in terms
of recognition rate on this dataset.
The second experiment is performed on Concordia University CENPARMI
handwritten numeral database. In this database, there are 600 samples for each of
ten digits (from 0 to 9), in which the frst 400 ones are used for training and the
rest for testing. Sample images of the digit zero from the CENPARMI database
are shown in Figure 3.4. In the experiment two kinds of original features: 256-di-
mensional Gabor transformation feature and 121-dimensional Legendre moment
feature extracted by Hu, Lou, Yang, Liu and Sun, (1999) are used respectively. The
best recognition rates using Foley-Sammon discriminant (FSD) (Foley and Sam-
mon, 1975), uncorrelated linear discriminant analysis (U-LDA), and LSVM, as well
as the accuracy using MNMSE are shown in Table 3.5. Note that the results with
asterisk are from (Yang, 2002). The accuracy of MNMSE is superior to all of face
recognition techniques except for LSVM.
The third experiment is performed on seven datasets available from the UCI
machine learning repository (Murphy & Aha, 1992). The recognition rates and
consumed time using LSVM and MNMSE are listed in Table 3.6. Statistical tests
(t-test and rank-sum test) both show that there is no signifcant difference between
is prohibited.
Table 3.5. Comparison of the accuracies of MNMSE with those of FSD, U-LDA,
and LSVM using CENPARMI database
FSD U-LDA LSVM MNMSE
Gabor feature 54.3%
*
84.7%
*
86.9% 85.8%
Legendre feature 51.7%
*
90.3%
*
93.35% 92.1%
Table 3.6. Comparison of the recognition rates (%) and the consumed time (sec) of
MNMSE with those of LSVM on seven machine learning datasets
Dataset LSVM MNMSE
name dimension Rate time Rate time
Heart 27013 83.70 8940 83.70 1.42
Ionosphere 35134 87.46 21.3 86.61 4.30
Iris 1504 98.00 1.06 98.00 4.55
Liver disorder 3456 70.14 772 62.90 1.52
Sonar 20860 76.44 5.06 73.08 3.97
Tic Tac Toe 9589 65.34 107 61.27 10.0
Wine 17813 94.38 2300 97.19 1.08
Figure 3.4 Sample images of the digit zero from the CENPARMI database
the recognition rates of the two methods at a signifcance level of 0.05. But there
does exit a signifcant difference between the times consuming by the two methods.
Considering the fact that the MNMSE is realized by the plain Matlab language
whereas the LSVM was realized by a special SVM tool (Ma, Zhao, & Ahalt, 2001)
the difference is amazing.
is prohibited.
3.4 mAxImum scAtter dI fference cl Ass If Ier
By viewing Fisher discriminant criterion as a special realization of a multi- objective
programming problem, another realization of the same multi-objective program-
ming problem leads to a novel classifcation-oriented linear discriminant criterion
called maximum scatter difference (MSD) (Song, Cheng, Yang, & Liu, 2004; Song,
Zhang, Chen, & Wang, 2007).
Between-class scatter
T
b
T
S w w
w w

and within-class scatter
T
w
T
S w w
w w

are two useful measures of the separability of projected samples. The larger the
score of between-class scatter is, the more separable the projected samples are;
the smaller the score of within-class scatter is, the more separable the projected
samples are.
Thus, the problem of fnding the optimal projection direction can be formulated
as a multi-objective programming problem:
Maximize:
T
b
T
S w w
w w
, (3.26)
Minimize:
T
w
T
S w w
w w
. (3.27)
There are mainly two ways to translate a multi-objective programming problem
into a single-objective programming problem. The frst is the goal programming
approach in which one of the objectives is optimized while the remaining objectives
are converted into constraints. The second is the combining objective approach in
which all objectives are combined into one scalar objective (Eschenauer, Koski,
& Osyczka, 1990).
Though the goal programming approach may also lead to a new linear dis-
criminant criterion, we will concern ourselves with the combining objective one
here. There exist two different principles to combine objectives. The frst is mul-
tiplicative and the second is additive. By applying the multiplicative principle to
is prohibited.
the multi-objective programming problem one can obtain the Fisher discriminant
criterion. If the additive principle is used the maximum scatter difference (MSD)
discriminant criterion is followed.
Let J
M
(w) denote
( )
T
b w
T
S C S - w w
w w
.
The optimization model corresponding to MSD discriminant criterion is as
follow:
max ( )
M
J w
, (3.28)
where the parameter C is a nonnegative constant which balances the relative merits
of the frst objective (3.26) to the second objective (3.27). The physical meaning of
the parameter will be discussed in detail in Subsection 3.4.3.
The Lagrangian function of the quadratic programming problem (3.28) is:
( , ) ( ) ( 1) ( ) ( 1)
T T
M b w
L J S C S = - - = - - - w w w w w w w
.
(3.29)
Let
( ,
0
L
=
w
w
, then we have:
( )
b w
S C S - = w w. (3.30)
By substituting Eq. (3.30) into J
M
(w) we obtain ( )
M
J = w .
Thus, the optimal projection direction determined by MSD is the eigenvector of
the matrix ( )
b w
S C S - corresponding to the largest eigenvalue. The optimal weight
calculated by MSD discriminant criterion is called MSD weight and is denoted by
*
w . It is obvious that
*
w is a function of the parameter C.
Unlike Fisher discriminant criterion which is inapplicable to the case where S
W
is
singular and LMLP discriminant criterion which is applicable only to these special
cases, the optimal projection direction of MSD discriminant criterion always exists
no matter whether the matrix S
W
is singular or not.
3.4.2 Relations between MSD Classifer and Other Classifers
Relations between the MSD weight
*
w
and weights calculated by certain well known
classifers are revealed in this subsection.
is prohibited.
3.4.2.1 On Nonsingular Within-Class Scatter Matrix
Let
*
( ) c w be the optimal projection direction determined by MSD when the param-
eter C assumes the value c, and
* *
*
* *
( ) ( ) ( )
( ) ( ( ))
( ) ( )
T
b w
M
T
c S c S c
F c J c
c c
-
= =
w w
w
w w
(3.31)
be the separability measure of projected samples along the direction
*
( ) c w .
Since
0
( )
( ) max
T
b w
T
S c S
F c

-
=
w
w w
w w
,
F(c) is equal to the largest eigenvalue of the matrix
( )
b w
S c S -
according to the
extremum property of generalized Rayleigh quotient.
Theorem3.1 F(c) is a monotone decreasing function. Especially, F(c) will be a
strictly monotone decreasing function if the within-class scatter matrix S
w
is non-
singular and the limit of F(c) is negative infnity when c is approaching infnity.
Proof: Let
1 2
c c <
, and
i
w be the unit eigenvector of the matrix ( )
b w
S c S - corre-
sponding to the largest eigenvalue, i = 1, 2. It is obvious that
1 1 1 1 2 1 2
( ) ( ) ( )
T T
b w b w
F c S c S S c S = - - w w w w
2 2 2 2 1 2 2
( ) ( )
T T
b w w
S c S c c S = - + - w w w w
2 2 1 2 2
( ) ( ) .
T
w
F c c c S = + - w w
Since S
w
is semi-positive defnite, we have
2 2
0
T
w
S w w . Thus,
1 2
( ) ( ) F c F c
, i.e.,
F(c) is a monotone decreasing function.
Especially, when S
w
is non-singular, S
w
is positive defnite. Thus, for any unit
vector w, we always have
0
T
w w
S > w w , (3.32)
where
w
is the smallest eigenvalue of S
w
.
In this case,
1 2
( ) ( ) F c F c > , i.e., F(c) is a strictly monotone decreasing func-
tion.
is prohibited.
Let
b
denote the largest eigenvalue of the matrix
b
S . Then, for any unit vector
w, the following inequality always holds
T
b b
S w w
. (3.33)
Combining Eq. (3.32) and Eq. (3.33), we have
* * * *
( ) ( ) ( ) ( ) ( )
T T
b w b w
F c c S c c c S c c = - - w w w w
.
It is obvious that lim ( )
c
F c
= -.
Thus we complete the proof of Theorem 3.1.
The curve of the function F(c) is illustrated in Figure 3.5.
Lemma3.1 F(c) is a continuous real-value function of variable c (Cao, 1980).
Theorem3.2If S
W
is nonsingular, there exists a unique positive root
0
c
of the equa-
tion F(c) = 0 such that the unit eigenvector of the matrix
0
( )
b w
S c S -
corresponding
to the largest eigenvalue is the Fisher weight.
Proof:It is obvious that
0
(0) max 0
T
b
b
T
S
F

= = >
w
w w
w w
.
From the proof of Theorem 3.1, we know that
*
( ) 0 F c < when
* b
w
c > .
Since F(c) is a continuous function, there must exist a point
0
c in the interval
*
(0, ) c such that
0
( ) 0 F c = .
Considering that F(c) is a strictly monotone function, we know that the point
0
c
is unique.
F(c) F(c)
c c
Figure 3.5 The curve of the function F(c): (a) Sw is nonsingular; (b) Sw is singu-
lar
is prohibited.
From
0
( ) 0 F c = , i.e.,
*
0 0
( ) ( ) 0
b w
S c S c - = w
, one can easily conclude that
* *
0 0 0
( ) ( )
b w
S c c S c = w w
.
Since the matrix
b
S has only one positive eigenvalue,
*
0
( ) c w is just the Fisher
weight
F
w .
Theorem 3.2 implies that Fisher classifer is a special MSD classifer when the
within-class scatter matrix is nonsingular.
3.4.2.2 Asymptotic Property of MSD
Theorem3.3.If S
W
is singular, the weight
*
( ) C w determined by MSD is approach-
ing the LMLP weight w
L
when the parameter C is approaching infnity.
Proof: Let w
b
, w
c
denote the unit eigenvectors of matrices S
b
,

(S
b
C S
W
) corre-
sponding to the largest eigenvalues
b
,
c
respectively.
Since S
W
is a singular matrix there exists a nonzero unit vector w
0
such that S
W

w
0
= 0. Considering the fact that S
b
is a semi-positive matrix we have:
0 0 0 0 0 0
1
max ( ) 0
T T T T
c b w b w b
S C S S C S S
=
= - - =
w
w w w w w w w w .
(3.34)
From the meanings of
c
and w
c
the following equation is always true for any
positive real number C:
( )
b w c c c
S C S - = w w
. (3.35)
Using equality (3.35), inequality (3.34) and the meaning of
b
we obtain:
1 1 1
( )
T T T
c w c c b c c c b c b
S S S
C C C
= - w w w w w w
. (3.36)
Noticing that the matrix S
W
is also semi-positive we fnd:
0
T
c w c
S w w . (3.37)
By combining facts (3.36) and (3.37) we can simply conclude that
lim 0
T
c w c
C
S
= w w . (3.38)
is prohibited.
Theorem 3.3 implies that LMLP classifer is in fact an asymptotic formof the
MSD classifer when the within-class scatter matrix is singular.
Theorem3.4.If S
W
is nonsingular, the weight w
*
(C) determined by MSD is ap-
proaching the unit eigenvector of the within-class scatter matrix corresponding to
the smallest eigenvalue when the parameter C is approaching positive infnity.
Proof: Let w
0
denote the unit eigenvector of matrix S
w
corresponding to the smallest
eigenvalue
0
, and w
c
be the unit eigenvector of matrix (S
b

C S
w
) corresponding
to the largest eigenvalue. Then, for any unit vector w, we have:
( ) ( )
T T
c b w c b w
S C S S C S - - w w w w. (3.39)
Let v
c
=

w
0
w
c
. Thus, the following inequality holds:
0 0 0 0
( ) ( )( ) ( )
T T
c b w c b w
S C S S C S + - + - v w v w w w
. (3.40)
It can be simplifed as:
0 0 0 0 0
( 2 ) 2
T T T T T T
c w c c w c b c b c b c b b
C S S S S S S + - + v v v w w w w w w w w w
.
(3.41)
Here
b
is the largest eigenvalue of the matrix S
b
.
Since
0 0 0 w
S = w w
, the inequality (3.41) can be rewritten as:
0 0
2 2 /
T T
c w c c b
S C + v v v w
. (3.42)
Considering the fact
0 c c
= + w v w
and
0 0
1
T T
c c
= = w w w w
we have:
0
2
T T
c c c
= - v v v w
. (3.43)
Thus, by substituting the equation (3.43) into the inequality (3.41) the following
inequality holds for any positive number C:
0
2 /
T T
c w c c c b
S C - v v v v
. (3.44)
Since
0
is the smallest eigenvalue of the matrix S
w
, it means that:
0
lim
T T
c w c c c
C
S
= v v v v
. (3.45)
is prohibited.
From the defnition of v
c
we know that lim 0
c
C
= v , i.e.,
0
lim
c
C
= w w
. (3.46)
The binary linear classifer based on the weight w
0
is called the smallest within-
class scatter classifer in this book.
3.4.2.3 Maximum Scatter Difference with C = 0
If the parameter C assumes zero value the MSD weight will degenerate into the
eigenvector of the matrix S
b
corresponding to the largest eigenvalue. By the defnition of
S
b
in (3.10) we know that it has an unique positive eigenvalue
1 2 1 2
( ) ( )
T
- - m m m m . The
corresponding unit eigenvector is
1 2
1 2
b
-
=
-
m m
w
m m
. (3.47)
In this case MSD classifer assigns a unknown sample x with the same label as
the nearest mean sample (or centroid) m
i
, = 1,2. This is a binary minimum distance
algorithm with Euclidean distance. We call this classifer as Centroid classifer.
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
0 1.0E-6 1.0E-2 1.0E2 1.0E6
The Parameter C
T
h
e

R
e
c
o
g
n
i
t
i
o
n

R
a
t
e
c
0
Figure 3.6. A sample curve of r(c) when the within-class scatter matrix is nonsin-
gular
is prohibited.
3.4.3 discussion on the Parameter c
Now we discuss the physical meaning of the parameter C in some detail. Let r(c) be
the recognition rate obtained by MSD classifer for a certain recognition problem
when the parameter C assumes value c.
If the within-class scatter matrix is nonsingular one sample curve of the func-
tion r(c) is illustrated in Figure 3.6 (The plot is based on experiments performed on
Australian dataset available from the UCI machine learning repository (Murphy
& Aha, 1992).)
From Figure 3.6 we fnd that there is a pulse in the sample curve. When the
parameter C is small enough the recognition rate obtained by MSD classifer is the
same as the one of the Centroid classifer; when the parameter C is large enough
the recognition rate obtained by MSD classifer is equal to the recognition rate of
the smallest within-class scatter classifer. When the parameter C is approaching
c
0
the recognition rate obtained by MSD classifer is approaching the maximum
value the recognition rate of Fisher classifer. It implies that the recognition rate
obtained by MSD classifer is not robust on the parameter C when the within-class
scatter matrix is nonsingular.
If the within-class scatter matrix is singular one sample curve of the function
r(c) is illustrated in Figure 3.7 (The plot is based on experiments performed on ORL
face image dataset with the frst fve images of each person used for training and
the rest images used for test.)
Figure 3.7. A sample curve of r(c) when the within-class scatter matrix is singu-
lar
is prohibited.
From Fig 3.7 we fnd that r(c) is nearly a nondecreasing function. When the
parameter C is small enough the recognition rate obtained by MSD classifer is the
same as the one of the Centroid classifer; when the parameter C is large enough
the recognition rate obtained by MSD classifer is equal to the recognition rate of
LMLP classifer. It implies that the MSD classifer can be applied to small size
problem such as face recognition without the need of parameter tuning.
Let w
b
denote the unit eigenvector of matrix S
b
eigenvalue, w
e
be the unit eigenvector of matrix S
w
corresponding to the smallest
eigenvalue if S
w
is nonsingular, or the LMLP weight w
L
if S
w
is singular. Thus,
when the parameter C increases from zero to positive infnity MSD weight w
*
(C)
will swerve from w
b
to w
e
gradually as illustrated inFigure 3.8. Moreover, the end
point of Fisher weight w
F
is always in the arc if it exists, i.e., the within-class scatter
matrix S
w
is nonsingular. We can conclude that when S
w
is nonsingular the parameter
C should be close to c
0
(see Theorem 3.2) such that w
*
(C) is close to Fisher weight
w
F
. When S
w
is singular the parameter C can be any large number.
The performance of MSD classifer is evaluated on ORL, Yale, a subset of FE-
RET (Phillips, Moon, Rizvi, & Rauss, 2000; Yang, Yang, & Frangi, 2003), and a
subset of AR (Yang, Zhang, Frangi, & Yang, 2004; Martinez & Benavente, 1998)
face image databases. In comparison with MSD, experimental results of certain
linear classifers such as LSVM, LMLP, and Centroid are also included. In addi-
x
1
x
2
x
n
w
b
w
*
(C)
w
e
O
Figure 3.8. The parameter controls the tradeoff between the weights w
b
and w
e
is prohibited.
tion, experimental results of well known facial feature extraction methods such
as Eigenface, Fisherface, orthogonal complementary space (OCS) (Liu, Cheng,
& Yang, 1992), nullspace LDA (N-LDA) (Chen, Liao, Ko, Lin and Yu, 2000), and
direct LDA (D-LDA) (Yu & Yang, 2001) in combination with the nearest neighbor
classifer are also reported here.
The experiment conducted on the ORL dataset contains ten runs. In each of
the ten runs, fve images of each person are used for training and the remaining
fve for testing. The images of each person numbered 1 to 5, 2 to 6, , 10 to 4 are
used as training samples for the frst, second,, and the tenth run respectively.
The average recognition rates of various face recognition methods are reported.
In the experiment performed on the Yale face image database, the leave-one-out
evaluation strategy was adopted.
The subset of FERET face image database used in this experiment includes 1400
images of 200 individuals (each individual has seven images). This subset involves
variations in facial expression, illumination, and pose. In the experiment, the facial
portion of each original image was automatically cropped based on the location
of eyes and the cropped image was resized to 8080 pixels and preprocessed by
histogram equalization. In each of the seven runs, four images of each person are
used for training and the remaining three images for testing. The images of each
Figure 3.10. All images of one individual from the AR database
Figure 3.9 Sample images from FERET database
( a) (b) (c) (d) (e) (f) (g) (h) ( i)
(j) (k) (l) (m) (n) (o) (p) (q) (r)
(s) (t) (u) (v) (w) (x) (y) (z)
is prohibited.
person numbered 1 to 4, 2 to 5, , 7 to 3 are used as training samples for the frst,
second, , and seventh run respectively. Sample images of one individual from
the subset of FERET database are illustrated in Figure 3.9.
The AR face database contains over 4,000 color face images of 126 people (70
men and 56 women), including frontal views of faces with different facial expres-
sions, lighting conditions and occlusions. The pictures of most persons were taken
in two sessions (separated by two weeks). Each section contains 13 color images
and 119 individuals (65 men and 54 women) participated in both sessions. The im-
ages of these 119 individuals were selected and used in the experiment. All color
images have been transformed into gray images. Each original image was denoted
by a 768576 matrix and had 256 gray levels. Figure 3.10 shows all samples of one
individual from the AR database, where (a)(m) are from Session 1 and (n)(z) are
from Session 2. The details of the images are: (a) and (n), neutral expression; (b)
and (o), smile; (c) and (p), anger; (d) and (q), scream; (e) and (r), left light on; (f) and
(s), right light on; (g) and (t), all sides light on; (h) and (u), wearing sun glasses; (i)
and (v), wearing sun glasses and left light on; (j) and (w), wearing sun glasses and
right light on; (k) and (x), wearing scarf; (l) and (y) wearing scarf and left light on;
and (m) and (z), wearing scarf and right light on.
Only the full facial images were considered here (no attempt was made to
handle occluded face recognition in each session). All images are normalized with
a resolution of 5040 and pre-processed by histogram equalization. The normal-
izedimages of one person are shown in Figure 3.11, where Figs. 3.11a, 3.11b, 3.11c,
3.11d, 3.11e, 3.11e, 3.11f, and 3.11g are from Session 1, and Figs. 3.11n, 3.11o, 3.11p,
3.11q, 3.11r, 3.11s, and 3.11t are from Session 2.
(a) (b) (c) (d) (e) (f) (g)
(n) (o) (p) (q) (r) (s) (t)
Figure 3.11. Normalized images of one individual from the AR database
is prohibited.
In the experiment, seven images of each person in the frst session are used for
training and the seven images in the second session for testing. The symbol *
means that orthogonal complementary space method is not applicable to large face
image databases such as FERET in our experiments. It is too time consuming.
Since the dimensions of input space in all datasets are very high, in order to save
computing time, input spaces are compressed by PCA transform before feature
Table 3.7. Comparison of the accuracies of MSD classifer with those of other face
recognition techniques
Datasets
ORL Yale FERET AR
M
e
t
h
o
d
s
Centroid 90.25% 91.52% 55.90% 63.33%
LSVM 95.25% 90.91% 62.95% 66.19%
LMLP 94.95% 94.55% 62.19% 65.95%
MSD 95.15% 92.73% 62.98% 66.07%
Eigenface 93.25% 81.82% 55.71% 65.12%
Fisherface 91.00% 97.58% 27.55% 59.88%
OCS 95.90% 98.18% * 59.88%
N-LDA 96.25% 98.18% 52.10% 60.12%
Direct LDA 94.95% 98.79% 47.62% 63.45%
Figure 3.12. Sample recognition rate curves of MSD and LSVM classifers
is prohibited.
extraction or classifer training. The recognition rates of various face recognition
methods on datasets ORL, Yale, FERET, and OR are reported in Table 3.7. From
the experimental results we fnd that in comparison with state-of-the-art face rec-
ognition methods MSD is very promising.
To check the robustness of MSD classifer for SSS problems we compare the
recognition rate curve of MSD with that of LSVM on the ORL face image dataset
when the parameter C varies from 0.001 to 1000. Here, the training sample set
consists of the frst fve images of each individual, and test sample set consists of
the rest of the images. The experimental results are illustrated in Figure 3.12. As
a result, we can conclude that MSD classifer is rather robust and can achieve very
high recognition rate.
3.5 summAr y
With this chapter, we frst provide a brief introduction to linear classifcation and
then present three novel classifcation-oriented linear discriminant criteria. The frst
one is the large margin linear projection which makes full use of the characteristic
of the SSS problems. Theoretical analysis demonstrates that LMLP is a special
linear support vector machine. Therefore it inherits the advantages of LSVM. As
LMLP is especially developed for the SSS problems, it is not applicable to other
type of pattern recognition problems. The second one is the minimum norm mini-
mum squared-error criterion which is a modifcation of the minimum squared-error
discriminant criterion. MNMSE has a unique closed-form solution which avoids
suboptimal points of traditional gradient descent approaches for MSE. The third
one is the maximum scatter difference which is a modifcation of Fisher discrimi-
nant criterion. Neither like FDC which is only applicable to situations when the
within-class scatter matrix is nonsingular, nor like LMLP which is only applicable
to situations when the within-class scatter matrix is singular, MSD can be used in
both cases. Theory analysis shows that MSD is an extension of Fisher discriminant
criterion when the within-class scatter matrix is nonsingular, and its asymptotic
form is LMLP when the parameter approaches infnity and the within-class scatter
matrix is singular.
r eferences
Bartlett, M. S., Movellan, J. R., & Sejnowski, T. J. (2002). Face recognition by in-
dependent component analysis. IEEE Trans. Neural Networks, 13(6), 1450-1464.
is prohibited.
Bian, Z. Q., & Zhang, X. G. (2000). Pattern Recognition (in Chinese). Beijing:
Qinghua University Press.
Burges, C. J. C. (1998). A tutorial on support vector machines for pattern recogni-
tion. Data Mining and Knowledge Discovery, 2(2), 121-167.
Cao, Z. H. (1980). Eigenvalue problems (in Chinese). Shanghai: Shanghai Science
and Technology Press.
33(10), 1713-1726.
Cheng, Y. P. (1989). Matrix Theory (in Chinese). Xian: North-west Industry Uni-
versity Press.
John Wiley & Sons.
Eschenauer, H., Koski, J., & Osyczka, A. (1990). Multicriteria Design Optimization.
Berlin: Springer-Verlag.
Trans. Computer, 24(3), 281-289.
Hsu, C., & Lin, C. (2002). A comparison of methods for multiclass support vector
machines. IEEE Transaction on Neural Networks, 13(2), 415-425.
Hu, Z. S., Lou, Z., Yang, J.Y., Liu, K., & Sun, C. Y. (1999). Handwritten digit rec-
ognition based on multi-classifer combination. Chinese J. Comput. (in Chinese),
22 (4), 369-374.
Jin, Z., Yang, J. Y., Hu, Z. S., & Lou, Z. (2001). Face recognition based on the un-
correlated discriminant transformation. Pattern Recognition, 34(7), 1405-1416.
Liu, K., Cheng, Y. Q., & Yang, J. Y. (1992). An effcient algorithm for Foley-Sammon
optimal set of discriminant vectors by algebraic method. International Journal of
Pattern Recognition and Artifcial Intelligence, 6(5), 817-829.
Ma, J., Zhao, Y., & Ahalt., S. (2001). OSU SVM Classifer Matlab Toolbox (ver
3.00). From http: //www.eleceng.ohio-state.edu/~maj/osu_svm.
is prohibited.
Martinez, A. M., & Benavente, R. (1998). The AR Face Database. CVC Technical
Report, no. 24.
Murphy, P. M., & Aha, D. W. (1992). UCI repository of machine learning databases.
http://www.ics.uci.edu/~mlearn/MLRepository.html.
Phillips, P. J., Moon, H., Rizvi, S. A., & Rauss, P. J. (2000). The FERET evaluation
methodology for face recognition algorithms. IEEE Trans. Pattern Analysis and
Machine Intelligence, 20(10), 1090-1104.
Song, F. X., Cheng, K., Yang, J. Y., & Liu, S. H. (2004). Maximum scatter differ-
ence, large margin linear projection and support vector machines. Acta Automatica
Sinica (in Chinese), 30(6), 890-896.
Song, F. X., Yang, J. Y., & Liu, S. H. (2004). Large margin linear projection and
165-174.
Turk, M., & Pentland, A. (1991). Face recognition using eigenfaces. Proc. IEEE
Conf. Computer Vision and Pattern Recognition, (pp. 586-591).
Vapnik, V. (1995). The Nature of Statistical Learning Theory. New York: Springer-
Verlag.
Yang, J. (2002). Linear projection analysis: theory, algorithms, and application
in Feature Extraction (in Chinese). Unpublished doctoral dissertation, Nanjing
University of Science & Technology, Nanjing, China.
Yang, J., Yang, J. Y., & Frangi, A. F. (2003). Combined fsherfaces framework. Im-
age and Vision Computing, 21(12), 1037-1044.
Yang, J., Zhang, D., Frangi, A. F., & Yang, J. Y. (2004). Two-dimensional pca: a
new approach to appearance-based face representation and recognition. IEEE Trans.
Yang, M. H. (2002). Kernel Eigenfaces vs. kernel Fisherfaces: face recognition
using kernel methods. Proc. Fifth IEEE Intl Conf. Automatic Face and Gesture
Recognition, (pp. 215-220).
Yu, H., & Yang, J. (2001). A direct lda algorithm for high-dimensional datawith
is prohibited.
Chapter IV
Orthogonal Discriminant
Analysis Methods
Abstr Act
In this chapter, we frst give a brief introduction to Fisher linear discriminant, Foley-
Sammon discriminant, orthogonal component discriminant, and application strategies
for solving the SSS problems. We then present two novel orthogonal discriminant
analysis methods, orthogonalized Fisher discriminant and Fisher discriminant
with Schur decomposition. At last, we compare the performance of several main
orthogonal discriminant analysis methods under various SSS strategies.
4.1 Introduct Ion
4.1.1 f isher l inear discriminant
Fisher linear discriminant (FLD) (Duda, Hart, & Stork, 2001) operates by learning a
discriminant matrix which maps a d-dimensional input space into an r-dimensional
feature space by maximizing the multiple Fisher discriminant criterion.
Specifcally, a Fisher discriminant matrix is an optimal solution of the following
optimization model:
max ( )
d r
T
B
F
T
W R
W
W S W
J W
W S W
=
. (4.1)
Orthogonal Discriminant Analysis Methods 59
is prohibited.
Here
d r
W R

is an arbitrary matrix, and S

B
and S
W
are the between- and within-
class scatter matrices, and
A
is the determinant of a square matrix A.
The between-class scatter matrix S
B
and the within-class scatter matrix S
W
are
defned as follows,
1
( )( )
l
T
B i i i
i
S N
=
= - -
m m m m , (4.2)
and
1 1
( )( )
i
l l
T
W i i i
i i
S S
= =
= = - -

x
x m x m . (4.3)
Here N
i
and m
i
are respectively the number and the mean of samples from the
ith class
i
, m the mean of samples from all classes, and l the number of classes.
It has been proved that if S
W
is nonsingular, the matrix composed of unit ei-
genvectors of the matrix
1
W B
S S
-
corresponding to the frst r largest eigenvalues is
an optimal solution of the optimization model defned in Eq. (4.1) (Wilks, 1962).
The matrix
1
W B
S S
-
is the Fisher discriminant matrix commonly used in Fisher linear
discriminant.
Since the matrix
1
W B
S S
-
is usually asymmetric, Fisher discriminant vectors, i.e.
column vectors of the Fisher discriminant matrix are unnecessary orthogonal to
each other.
4.1.2 Foley-Sammon Discriminant
Many researchers think that it is helpful to eliminate linear dependencies among
discriminant vectors by making them orthogonal to each other. Foley-Sammon
discriminant (FSD) is a feature extraction method that does this using the optimal
discriminant vectors. Optimal discriminant vectors are derived from the multiple
Fisher discriminant criterion which is subject to orthogonality constraints initially
for binary classifcation tasks (Sammon, 1970; Foley & Sammon, 1975). In detail,
the frst discriminant vector of FSD is the frst Fisher discriminant vector, i.e. a unit
eigenvector of the matrix
1
W B
S S
-
corresponding to the largest eigenvalue. After the
frst k (1 k r < ) discriminant vectors
1,
...,
k
w w have been calculated, the (k+1)th
discriminant vector
1 k +
w of FSD is then one of the optimal solutions to the follow-
ing optimization model
0
0, 1,...,
max ( )
T
i
T
B
T
i k
W
S
J
S
= =
=
w w
w w
w
w w
. (4.4)
is prohibited.
Here, J
0
(w) is a special form of the multiple Fisher discriminant criterion func-
tion, J
F
(W) in the model (4.1) in the case of r = 1.
Duchene and Leclercq (1998) extended FSD to multiclass classifcation problems
and presented an algorithm for calculating FSD discriminant vectors. Subsequently,
Jin, Yang, Hu, and Lou (2001) further simplifed the calculation procedure. They
showed that a unit eigenvector of the matrix
1
W B
S PS
-
eigenvalue could be the (k+1)th discriminant vector of FSD. The matrix P is cal-
culated by the formula
1 1 1
( )
T T
W W
P I D DS D DS
- - -
= -
. (4.5)
Here I is an identity matrix, and
1
[ ,..., ]
T
k
D = w w .
4.1.3 orthogonal component discriminant
Previously studies demonstrated that FSD was quite effective in pattern recogni-
tion. Unfortunately, the calculation procedure for FSD discriminant vectors is too
complex and time- consuming. To overcome the drawbacks of FSD, many orthogo-
nal discriminant analysis methods have been proposed in succession. Orthogonal
component discriminant (OCD) (Sun and Wu, 2003) is one example of them.
Sun and Wu (2003) proposed OCD based on intuition. They used the following
discriminant criterion instead of the multiple Fisher discriminant criterion,
1
max ( )
d r
T
B W
S
T
W R
W
W S S W
J W
W S W
=
. (4.6)
It can be proved that the matrix composed of unit eigenvectors of the matrix
1 1
W B W
S S S
- -
corresponding to the frst r largest eigenvalues is an optimal solution of
the optimization model (4.6). The column vectors of this eigenvector matrix are
discriminant vectors of OCD. Since the matrix
1 1
W B W
S S S
- -
is symmetrical, there is
a set of orthogonal eigenvectors of the matrix. These orthogonal eigenvectors are
the discriminant vectors of OCD.
4.1.4 Application Strategies for Small Sample Size Problems
A SSS problem arises when the dimensionality of an input space is larger than the
number of training samples. In such cases, since the within-class scatter matrix S
W

is singular, it is not possible to properly calculate the discriminant vectors for FLD,
is prohibited.
FSD, OCD, or other standard LDA methods. As a result, these discriminant feature
extraction methods cannot be applied directly to SSS problems.
In order to make standard discriminant feature extraction methods applicable to
SSS problems with either no or only minor modifcation, it is necessary to adopt one
of application strategies for SSS problems. For convenience, we call these application
strategies for SSS problems as SSS strategies. A SSS strategy is a procedure in which
a high-dimensional input space is compressed into a low-dimensional intermedi-
ate space such that in which standard or extended feature extraction methods are
applicable. In the following we present three such SSS strategies.
4.1.4.1 Strategy One
SSS Strategy One is primarily used in combination with FLD in the famous facial
feature extraction methodFisherface (Belhumeur, Hespanha, & Kriengman, 1997;
Swets & Weng, 1996). This strategy frst uses principal component analysis (PCA)
transformation to compress a high-dimensional input space into the range of the total
scatter matrix (the sum of the between-class scatter matrix and the within-class
scatter matrix) of training samples. It then selects the largest n (n N l - ) principal
components of the samples. Here, N is the number of training samples, and l is the
number of classes. Finally, standard feature extraction methods such as FLD, FSD,
and OCD are performed on this n-dimensional subspace. Actually, Fisherface is
SSS Strategy One + FLD.
Strategy One has widely been criticized for two drawbacks. First, it might lose
some useful discriminant information by abandoning some minor components in
the PCA stage. Second, theoretically it can not guarantee the non-singularity of
the within-class scatter matrix. A similar SSS strategy has been proposed by Liu
and Wechsler (2000). It is also believed to only partially solve the SSS problem and
fails to obviate the risk of losing useful discriminant information.
4.1.4.2 Strategy Two
Strategy Two frst compresses a high-dimensional input space into the range of the
total scatter matrix of training samples by PCA transformation. Then, extended ver-
sions of standard feature extraction methods are applied on this reduced space.
Extended Fisher linear discriminant (EFLD) is proposed as a substitute for
FLD (Fukunaga, 1990). The key idea is to replace the within-class scatter matrix
S
W
in FLD with the total scatter matrix S
T
. Liu, Cheng, Yang and Liu (1992) had
showed that EFLD was equivalent to FLD if the matrix S
W
was nonsingular. The
idea of EFLD has been extensively investigated in pattern recognition literature
is prohibited.
(Fukunaga, 1990; Liu, Cheng, Yang, & Liu, 1992; Guo, Li, Yang, Shu, & Wu, 2003;
Jing, Zhang, & Tang, 2004).
The extended multiple Fisher discriminant criterion is defned as,
max ( )
d r
T
B
E
T
W R
T
W S W
J W
W S W
=
. (4.7)
Here S
T =
S
B
+

S
W
is the total scatter matrix. The frst r discriminant vectors of
EFLD are unit eigenvectors of the matrix
1
T B
S S
-
corresponding to the frst r largest
eigenvalues. Similarly, we can extend FSD and OCD to extended FSD (EFSD) and
extended OCD (EOCD), by replacing S
W
with S
T
in calculation procedures for FSD
and OCD discriminant vectors.
Discriminant vectors of EFSD should meet the following requirements. The
frst discriminant vector of EFSD is a unit eigenvector corresponding to the larg-
est eigenvalue of the matrix
1
T B
S S
-
. After the frst k (1 k r < ) discriminant vectors,
1,
...,
k
w w
have been calculated, the (k+1)th discriminant vector
1 k +
w of EFSD is a
unit eigenvector of the matrix
1
T B
S PS
-
corresponding to the largest eigenvalue. The
matrix P is calculated by the formula
1 1 1
( )
T T
T T
P I D DS D DS
- - -
= -
. (4.8)
Here I is again an identity matrix, and
1, 2
[ ,..., ]
T
k
D = w w w . The frst r discrimi-
nant vectors of EOCD are unit eigenvectors of
1 1
T B T
S S S
- -
corresponding to the frst
r largest eigenvalues.
Although EFLD, EFSD, and EOCD cannot be directly applied to SSS problems,
they can safely be performed in the range of S
T
. That is, SSS problems can be
completely solved in theory if these extended feature extraction methods are used
in combination with a PCA transformation. It follows the SSS Strategy Two.
4.1.4.3 Strategy Three
Whereas the frst two SSS strategies both involve a Karhunen-Loeve expansion
based on the total scatter matrix, Strategy Three uses another kind of Karhunen-
Loeve expansion. SSS Strategy Three frst compresses a high-dimensional input
space into the range of the within-class scatter matrix of training samples by a
Karhunen-Loeve expansion based on the within-class scatter matrix. Then, the standard
feature extraction methods such as FLD, FSD, and OCD are safely performed on
this reduced space.
The procedure for calculating discriminant vectors of FLD, FSD, and OCD
under Strategy Three is as follows.
is prohibited.
Step1. Perform eigenanalysis on the matrix S
W
through singular value decomposi-
tion theorem (Golub & Loan, 1996). Let
1 2
[ , ,..., ]
n
P = p p p
be the matrix composed
of unit eigenvectors of S
W
corresponding to nonzero eigenvalues. Here
( )
W
n rank S =

is the rank of S
W
.
Step2. Compress the input space into the range of S
W
by projecting an input vector
x to P
T
x.
Step3. Calculate the between- and within-class scatter matrices of projected samples
by using formulae
T
B B
S P S P =
and
T
W W
S P S P =
.
Step4. Calculate the discriminant matrix V based on
B
S
and
W
S
using FLD, FSD,

or OCD.
Step5. Let W = PV. The column vectors of W are discriminant vectors what we
seek.
4.2 orthogon Al Ized fI sher dI scr ImInAnt
Although orthogonal component discriminant successfully decreases the compu-
tational complexity of Foley-Sammon discriminant, the theoretical background
of OCD is poor. There is no convincing explanation for its high accuracy. In this
section, a more convincing orthogonal discriminant feature extraction method
orthogonalized Fisher discriminant (OFD) is presented.
Suppose
1
,...,
r
w w are r Fisher discriminant vectors and v
1
is w
1
. Assume k vectors
1
,...,
k
v v
(1 1 k r - ) have been obtained. The (k+1)th OFD vectors
1 k +
v is calcu-
lated as follows:
1
1 1
1
T
k
i k
k k i
T
i
i i
+
+ +
=
= -
v w
v w v
v v
. (4.9)
It is easy to verify that discriminant vectors
1 1
,..., ,
k k +
v v v are orthogonal to
each other.
is prohibited.
We can rewrite Eq. (4.9) as
1
1
k
k k ik i
i
-
=
= +
v w w, here ik,
1,..., 1, i k = -

2,..., k =
r
are real numbers. Let
1
[ ,..., ]
r
V = v v , then we have V W = , where is an upper
triangular matrix with main diagonal elements being all one. Thus,
( )
( ) ( ) ( )
( )
T T T T T T
B B B
F F F
T T T T T T
W W W
W S W W S W W S W
J V J W J W
W S W W S W W S W

= = = = =

(4.10)
It means that the matrix which consists of OFD discriminant vectors is also an
optimal solution to the optimization model (4.1).
To evaluate the performance of OFD eight datasets are used in this subsection.
The frst is a subset of the FERET face database (please refer to Section 3.4.4). The
used subset contains 700 images from 100 individuals (each person has 7 different
images). Each image was preprocessed using the same approach shown in Section
3.4.4. For each person three images are randomly selected as training samples
and the remaining four are used for testing. Thus the total amount of training and
testing samples are 300 and 400 respectively. The accuracy is estimated by using
a ten-run average.
The second dataset is the CENPARMI handwritten numeral database (please
refer to Section 3.3.3). This experiment uses the same training set and the same
testing test as the second experiment in Section 3.3.3. The 256-dimensional Gabor
transformation feature and the 121-dimensional Legendre moment feature are also
used in the experiment. The third dataset is the NUST603 handwritten Chinese
character database built in Nanjing University of Science and Technology. The
database contains 19 groups of Chinese characters, which are collected from bank
checks. There are 400 samples for each character (7600 in total), in which the frst
200 ones are used for training and the others for testing. Sample images from the
NUST603 handwritten Chinese character database are showed in Figure 4.1. Two
kinds of features: 128-dimensionl cross feature and 128-dimensional peripheral
feature are used in the experiment. The other fve datasets are from the UCI Ma-
chine Learning Repository (Murphy & Aha, 1992). Accuracies are estimated by
using the leave-one-out strategy.
is prohibited.
In the experiment on the FERET face image database SSS Strategy One was used.
In experiments performed on all datasets except for CENPARMI and NUST603,
the nearest neighbor classifer is used. In experiments on the CENPARMI hand-
written digit database and the NUST603 handwritten Chinese character database,
the quadratic Bayesian classifer is used. The performance of FSD and OFD on
the FERET database is shown in Table 4.1. The accuracies of the FSD and OFD
on the CENPARMI handwritten numeral database are shown in Table 4.2. The
experimental results of the FSD and OFD on the NUST603 handwritten Chinese
character database and the fve UCI Machine Learning datasets are shown in Table
4.3 and Table 4.4, respectively. These experimental results demonstrate that OFD
is more accurate and much more effcient than FSD.
4.3 FiShER DiSCRiMiNANt With SChUR DECOMPOSiti ON
In this section, another promising orthogonal discriminant feature extraction
methodFisher discriminant with Schur decomposition (FDS) is provided.
Figure 4.1. Sample images from the NUST603 handwritten Chinese character
database
Table 4.1. Comparison of performance of FSD and OFD on the FERET database
Average time for feature extraction and
pattern classifcation (sec)
Average accuracy (%)
FSD OFD FSD OFD
55.88 12.84 66.43 67.11
is prohibited.
Table 4.2. Accuracies and total times for feature extraction on the CENPARMI
database
Number of extracted
features
Gabor feature Legendre feature
FSD (%) OFD (%) FSD (%) OFD (%)
1
3
6
9
28.05
39.25
43.05
49.35
28.05
45.50
73.95
80.75
30.40
35.05
39.35
43.90
30.40
64.65
84.45
89.35
Time (sec) 9.5940 1.7970 1.0470 0.4530
Table 4.3. Accuracies and total times for feature extraction on the NUST603 da-
tabase
Dataset Dimension
No. of
classes
No. of
features
FSD OFD
Accuracy
(%)
Time
(sec)
Accuracy
(%)
Time
(sec)
Glass 2149 6 5 54.67 2.500 68.69 1.860
Iris 1504 3 2 97.33 0.907 96.00 0.844
Vehicle 84618 4 3 70.33 28.187 74.82 28.422
Wine 17813 3 2 92.13 1.344 98.88 1.328
Zoo 10116 7 6 96.04 2.843 96.04 2.563
Table 4.4. Performance comparison of OFD and FSD on fve UCI datasets
Number of extracted
features
Cross feature Peripheral feature
FSD (%) OFD (%) FSD (%) OFD (%)
1
6
12
18
30.24
46.34
76.18
81.92
30.24
85.92
93.03
93.13
23.63
81.53
90.61
91.03
23.63
90.16
95.21
96.05
Time (sec) 12.938 12.359 14.078 12.796
It is well-known that Schur decomposition is a natural extension of and good
substitution for eigenanalysis when the matrix indented to be analyzed is asymmetric
(Golub and Loan, 1996). Let A be a real square matrix, its Schur decomposition
is
T
A UTU = . Here U is an orthogonal matrix, and T is a pseudo upper diagonal
is prohibited.
matrix with the real eigenvalues of the matrix A on the diagonal and the complex
eigenvalues in 2-by-2 blocks on the diagonal.
Instead of performing eigenanalysis on the matrix
1
W B
S S
-
in FLD, Schur decom-
position is carried out on
1
W B
S S
-
in FDS. Let the Schur decomposition of
1
W B
S S
-
be
denoted by
T
UTU , and
1
,...,
d
u u be all column vectors of the matrix U, i.e. Schur
vectors of
1
W B
S S
-
. It is obvious that
1
,...,
d
u u are orthogonal to each other. Assume
1
,...,
r
u u
to be Schur vectors of
1
W B
S S
-
corresponding to the frst r largest eigenvalues,
i.e. the frst r largest diagonal elements of the matrix T. Thus, we can compress a
high-dimensional input space R
d
into a low-dimensional feature space R
r
by a map-
ping,
T
V x x . Here
1
[ ,..., ]
r
V = u u is the discriminant matrix of FDS, and
1
,...,
r
u u
are FDS discriminant vectors.
Similar to FLD and OFD, the discriminant matrix of FDS is also an optimal
solution to the optimization model (4.1). The following theorem reveals the fact.
Theorem4.1 Let
1 2
, ,...,
r
u u u be discriminant vectors of FDS. Thus, we have
1 2 1 2
1
1 2 1 2
[ , ,..., ] [ , ,..., ]
([ ,..., ]) max ( )
[ , ,..., ] [ , ,..., ]
d r
T
r B r
F r F
T
W R
r W r
S
J J W
S

= =
u u u u u u
u u
u u u u u u
. (4.11)
Proof.Since
1 2
, ,...,
r
u u u are Schur vectors of the matrix
1
W B
S S
-
corresponding to the
frst r largest eigenvalues, we have
1
W B j j j
S S
-
= u u , 1,..., j r = . (4.12)
Here
j
is the jth largest eigenvalue of the matrix
1
W B
S S
-
.
From the formula (4.12) it follows that
1 1 1 1 1
[ ,..., ] [ ,..., ] ( ,..., )[ ,..., ] [ ,..., ]
T T
r B r r r W r
S diag S = u u u u u u u u
.

(4.13)
Thus, we have
1 1
1
1
1 1
[ ,..., ] [ ,..., ]
([ ,..., ]) max ( )
[ ,..., ] [ ,..., ]
d r
T
r
r B r
F r j F
T
W R
j
r W r
S
J J W
S

=
= = =
u u u u
u u
u u u u
.

(4.14)
is prohibited.
The performance of FDS is evaluated using two benchmark datasets: the ORL
(please refer to Section 3.2.3 ) and AR (please refer to Section 3.4.4) face image
databases. For the ORL database, ten tests are conducted on this database for each
dimensionality reduction method. In each test fve images of each individual are
used for training and the remaining fve for testing. Thus the total amount of train-
ing samples and testing samples are both 200. The training samples are taken in
turn. That is, the images of each person numbered 1 to 5, 2 to 6, , 10 to 4 are
used as training samples respectively. For the AR database, 1666 images from 119
individuals are used for experiments (each has fourteen different images with seven
taken in the frst session and seven taken in the second session). All images are also
normalized and preprocessed using histogram equalization as shown in Section
3.4.4. In the experiment, seven images of each person in the frst session are used
for training and the seven images in the second session for testing.
In all experiments in this subsection, SSS Strategy One is used for FLD, FSD,
and FDS. As usual, the nearest neighbor classifer with Euclidean distance is used.
In order to evaluate the performance of FDS properly, experimental results of
N-LDA proposed by Chen, Liao, Ko, Lin and Yu (2000) are also included. It is
interesting that discriminant vectors of N-LDA are also orthogonal to each other.
5 10 15 20 25 30 35
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Number of extracted features
A
v
e
r
a
g
e
r
e
c
o
g
n
i
t
i
o
n
r
a
t
e
FLD
FSD
PCA+N-LDA
FDS
5 10 15 20 25 30 35
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
A
v
e
r
a
g
e
r
e
c
o
g
n
i
t
i
o
n
r
a
t
e
FLD
FSD
PCA+N-LDA
FDS
Figure 4.2. Average recognition rates of FLD, FSD, N-LDA, and FDS vs. the number
of extracted features on the ORL face image database
is prohibited.
Although N-LDA can directly compress a high-dimensional input space in theory,
its memory- and time-cost are extremely high. As a rule N-LDA is usually used in
combination with a PCA transformation in face recognition, i.e. PCA + N-LDA.
Figure 4.2 displays trends of average recognition rates of various dimensionality
reduction methods on the ORL face image database when the number of extracted
features increases. We fnd that FDS is as accurate as FSD, better than FLD, and
competes with N-LDA when the number of extracted features is large enough.
Figure 4.3 shows trends of recognition rate of various dimensionality reduction
methods on the AR face image database when the number of extracted features
is increasing. We fnd that FDS is the most accurate technique for feature dimen-
sionality reduction.
Table 4.5 lists average times consumed by various dimensionality reduction
methods and the corresponding standard deviations on the ORL database when 39
features are extracted. From Table 4.5 we fnd that FDS is near the most effcient
dimensionality reduction method N-LDA, and much more effcient than FSD. Table
4.6 lists times consumed by various dimensionality reduction techniques on the
AR database when 119 features are extracted. From Table 4.6 we fnd that FDS is
the most effcient dimensionality reduction method and much more effcient than
FSD. It should be pointed out that the poor effciency of FSD is mainly due to its
iterative calculation procedure for discriminant vectors.
Figure 4.3. Recognition rates of FLD, FSD, N-LDA, and FDS vs. the number of
extracted features on the AR face image database
54 64 74 84 94 104 114
0.4
0.45
0.5
0.55
0.6
0.65
Number of extracted f eatures
R
e
c
o
g
n
i
t
i
o
n
r
a
t
e
FLD
FSD
PCA+N-LDA
FDS
is prohibited.
4.4 comPAr Ison of orthogon Al dI scr ImInAnt
AnAl ys Is methods
The performance of several typical orthogonal discriminant analysis methods, i.e.
FSD, OCD, OFD, and FDS is evaluated using the three SSS strategies on the ORL
and FERET face image databases. For comparison, we also present the experimental
results of FLD. In all experiments the nearest neighbor classifer with Euclidean
distance is used.
4.4.1 Database and Experimental Settings
The ORL (please refer to Section 3.2.3) is a typical small-scale uniform face image
database. We conducted ten runs on this dataset for each combination of a feature
extraction method and an SSS strategy. In each runs fve images of each person
were used for training and the remaining fve for testing. This made a total of 200
training and 200 testing samples. The training samples are taken in turn. That is,
the images of each person numbered 1 to 5, 2 to 6, , 10 to 4 are used as training
samples.
The FERET (please refer to Section 3.4.4) is a typical large-scale noisy face image
database. In this section the same subset of FERET including 1400 images of 200
individuals as shown in Section 3.4.4 was used Each image was also preprocessed
Table 4.5. Average times and standard deviations of various methods on the ORL
database when 39 features are extracted (sec.)

Dimensionality Reduction Methods
FLD FSD N-LDA FDS
mean 0.8124 7.2626 0.5502 0.7608
std 0. 0321 0. 0.0732 0. 0178 0. 0129
Table 4.6. Times spent for dimensionality reduction of various methods on the AR
database when 119 features are extracted (sec.)
Dimensionality Reduction Methods
FLD FSD N-LDA FDS
Time 23.312 1916.2 58.828 16.765
is prohibited.
Figure 4.5. Average recognition rates of FLD, FSD, OCD, OFD, and FDS under SSS
Strategy Two vs. the number of extracted features on two face image databases
Strategy One vs. the number of extracted features on two face image databases
Strategy One vs. the number of extracted features on two face image databases
using the same approach shown in Section 3.4.4. Three images of each individual
were randomly chosen for training and the remaining four images were used for
testing. Thus, the training sample set size is 600 and the testing sample set size is
800. We ran the system ten times and obtained ten different training and testing
sample sets.
7 15 23 31 39
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
A
v
e
r
a
g
e
r
e
c
o
g
n
i
t
i
o
n
r
a
t
e
FLD1
FSD1
OCD1
OFD1
FDS1
39 79 119 159 199
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
A
v
e
r
a
g
e
r
e
c
o
g
n
i
t
i
o
n
r
a
t
e
FLD1
FSD1
OCD1
OFD1
FDS1
(a) ORL (b) FERET
7 15 23 31 39
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
FLD2
FS D2
OCD2
OFD2
FDS 2
39 79 119 159 199
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
FLD2
FS D2
OCD2
OFD2
FDS 2
(a) ORL (b) FERET
is prohibited.
4.4.2 Accuracy of Various Methods under the SSS Strategy
one
Figure 4.4 shows error bars of recognition rates of various feature extraction methods
under the SSS Strategy One on the ORL and FERET face image databases with
varying number of extracted features. Here, FLD1, FSD1, OCD1, OFD1, and FDS1
stand for FLD, FSD, OCD, OFD, and FDS in combination with SSS Strategy One
respectively. In fact, FLD1 is the well-known Fisherface.
t wo
Figure 4.5 shows error bars of recognition rates of various feature extraction meth-
ods under the SSS Strategy Two on the ORL and FERET face image databases
with varying number of extracted features. From Figure 4.5 we can conclude that
FLD is signifcantly inferior to all orthogonal discriminant analysis methods except
OCD.
t hree
Figure 4.6 shows error bars of recognition rates of various feature extraction methods
under the SSS Strategy Three on the ORL and FERET face image databases with
varying number of extracted features. From Figure 4.6 we fnd that FLD is almost
as effective as OCD, OFD, and FDS, and signifcantly more effective than FSD.
Strategy Three vs. the number of extracted features on two face image databases
7 15 23 31 39
0.4
0.5
0.6
0.7
0.8
0.9
1
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
FLD3
FS D3
OCD3
OFD3
FDS 3

39 79 119 159 199
0. 45
0.5
0. 55
0.6
0. 65
0.7
0. 75
0.8
0. 85
0.9
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
FLD3
FS D3
OCD3
OFD3
FDS 3
is prohibited.
Figure 4.7. Average times for feature extraction of various methods under the three
SSS strategies on two databases when 39 or 199 features are extracted
Figure 4.8. Average recognition rates of FLD under SSS Strategy One, Two, and
Three vs. the number of extracted features on two face image databases
4.4.5 Effciency Comparison of Various Methods
Figure 4.7 shows the average times spent for feature extraction of various methods
under the three SSS strategies on the ORL and FERET databases when 39 or 199
features are extracted. The time unit is second. From Figure 4.7, we see that FDS,
FLD, OCD, and OFD are similar in effciency, and they are much more effcient
than FSD no matter what kind of SSS strategy is used.
One Two Three
0
2
4
6
8
10
12
14
16
Application s trategy for S 3 problems
A
v
e
r
a
g
e

f
e
a
t
u
r
e

e
x
t
r
a
c
t
i
o
n

t
i
m
e

FLD
FS D
OCD
OFD
FDS

One Two Three
0
10
20
30
40
50
60
70
Application s trategy for S 3 problems
T
i
m
e

f
o
r

f
e
a
t
u
r
e

e
x
t
r
a
c
t
i
o
n

(
s
e
c
)
FLD
FS D
OCD
OFD
FDS
(a) ORL (b) FERET
7 15 23 31 39
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
FLD1
FLD2
FLD3

39 79 119 159 199
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
FLD1
FLD2
FLD3
(a) ORL (b) FERET
is prohibited.
Figure 4.11. Average recognition rates of OFD under SSS Strategy One, Two, and
Figure 4.9. Average recognition rates of FSD under SSS Strategy One, Two, and

7 15 23 31 39
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
FS D1
FS D2
FS D3

39 79 119 159 199
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
FS D1
FS D2
FS D3
(a) ORL (b) FERET
Figure 4.10. Average recognition rates of OCD under SSS Strategy One, Two, and

7 15 23 31 39
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
OCD1
OCD2
OCD3
39 79 119 159 199
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
OCD1
OCD2
OCD3
(a) ORL (b) FERET
7 15 23 31 39
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
OFD1
OFD2
OFD3

39 79 119 159 199
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
OFD1
OFD2
OFD3
(a) ORL (b) FERET
is prohibited.
Figure 4.12. Average recognition rates of FDS under SSS Strategy One, Two, and

7 15 23 31 39
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
FDS 1
FDS 2
FDS 3

39 79 119 159 199
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
FDS 1
FDS 2
FDS 3
(a) ORL (b) FERET
4.4.6 discussions on sss strategies
It is generally believed that since SSS Strategy Two holds all discriminant informa-
tion it is defnitely superior to the other two SSS Strategies. But our experimental
results demonstrate that it is not always true.
Figure 4.8-4.12 shows curves of average recognition rates of FLD, FSD, OCD,
OFD, and FDS under the SSS Strategy One, Two, and Three on the ORL and FERET
face image databases with the varying number of extracted features. From Figure
4.8-4.12 we fnd that SSS Strategy Three is the most effective and SSS Strategy
Two is the least effective in general.
4.5 summAr y
In this chapter we frst briefy review orthogonal discriminant analysis and its
related application strategies to small sample size problems. Then we present two
novel orthogonal discriminant analysis methods: one is orthogonalized Fisher
discriminant; the other is Fisher discriminant with Schur decomposition. Theoreti-
cal analysis and experimental studies showed that OFD and FDS are all optimal
solutions to multiple Fisher discriminant criterion, and they can compete with
FSD in accuracy and more effcient than the latter. At last we thoroughly compare
the performance of several main orthogonal discriminant analysis methods under
various SSS strategies.
is prohibited.
r eferences
33(10), 1713-1726.
principal component analysis. IEEE Trans. Pattern Anal. Machine Intell., 10(6),
978-983.
John Wiley & Sons.
Trans. Comput., 24(3), 281-289.
Fukunaga, K. (1990). Introduction to Statistical Pattern Recognition. New York:
Academic.
Golub, G. H., & Loan, C. F. V. (1996). Matrix Computations. Baltimore and London:
The Johns Hopkins University Press.
Guo, Y. F., Li, S. J., Yang, J. Y., Shu, T. T., & Wu, L. D. (2003). A generalized
Foley-Sammon transform based on generalized Fisher discriminant criterion and
its application to face recognition. Pattern Recognition Letters, 24(1-3), 147-158.
Jin, Z., Yang, J. Y., Hu, Z. S., & Lou, Z. (2001). Face recognition based on uncor-
Jing, X. Y., Zhang, D., Tang, Y. Y. (2004). An improved LDA approach. IEEE Transac-
tions on Systems, Man, and Cybernetics, Part B: Cybernetics, 34(5), 1942-1951.
Liu , C. J., & Wechsler, H. (2000). Robust coding schemes for indexing and retrieval
from large face databases. IEEE Trans. Image Processing, 9(1), 132-137.
Liu, K., Cheng, Y. Q. Yang, J. Y., & Liu, X. (1992). An effcient algorithm for Foley-
Sammon optimal set of discriminant vectors by algebraic method. International
Murphy, P. M., & Aha, D. W. (1992). UCI repository of machine learning databases.
http://www.ics.uci.edu/~mlearn/MLRepository.html.
is prohibited.
Sammon, J.W. (1970). An optimal discriminant plane. IEEE Trans. Comput., C-19,
826-829.
Sun, D. R., & Wu, L. N. (2003). Face recognition using orthonormal discriminant
subspace method. Proceedings of the International Conference on Active Media
Technology, (pp.314-319).
Swets, D. L., & Weng, J. (1996). Using discriminant eigenfeatures for image retrieval.
IEEE Trans. Pattern Anal. Machine Intell., 18(8), 831-836.
Wilks, S. S. (1962). Mathematical Statistics. New York: Wiley.

is prohibited.
Chapter V
Parameterized Discriminant
Analysis Methods
Abstr Act
In this chapter, we mainly present three kinds of weighted LDA methods. In Sections
5.1, 5.2 and 5.3, we respectively present parameterized direct linear discriminant
analysis, weighted nullspace linear discriminant analysis and weighted LDA in
the range of within-class scatter matrix. We offer a brief summery of the chapter
in Section 5.4.
5.1 PAr Ameter Ized dI rect lI ne Ar dI scr ImInAnt
AnAl ys Is
5.1.1 introduction
Direct LDA (D-LDA) (Yu & Yang, 2001) is an important feature extraction method
for SSS problems. It frst maps samples into the range of the between-class scatter
matrix, and then transforms these projections using a series of regulating matrices.
D-LDA can effciently extract features directly from a high-dimensional input space
without the need to frst apply other dimensionality reduction techniques such as
PCA transformations in Fisherfaces (Belhumeur, Hespanha, & Kriengman, 1997)
or pixel grouping in nullspace LDA (N-LDA) (Chen, Liao, Ko, Lin, & Yu, 2000),
and as a result has aroused the interest of many researchers in the feld of pattern
recognition and computer vision. Indeed, there are now many extensions of D-LDA,
Parameterized Discriminant Analysis Methods 79
is prohibited.
such as fractional D-LDA (Lu, Plataniotis, & Venetsanopoulos, 2003a), regularized
D-LDA (Lu, Plataniotis, &Venetsanopoulos, 2003b; Lu, Plataniotis, & Venetsano-
poulos, 2005), kernel D-LDA (Lu, Plataniotis, & Venetsanopoulos, 2003c), and
boosting D-LDA (Lu, Plataniotis, Venetsanopoulos, & Li, 2006).
But there nonetheless remain some questions as to its usefulness as a facial feature
extraction method. First, as been pointed out in Lu, Plataniotis and Venetsanopoulos
(2003b; Lu, Plataniotis, & Venetsanopoulos, 2005), D-LDA performs badly when
only two or three samples per individual are used. Second, regulating matrices in
D-LDA are either redundant or probably harmful. The second drawback of D-LDA
has not been seriously addressed in previous studies.
In this section, we present a new feature extraction methodparameterized direct
linear discriminant analysis (PD-LDA) for SSS problems (Song, Zhang, Wang, Liu,
& Tao, 2007). As an improvement of D-LDA, PD-LDA inherits advantages of D-
LDA such as direct and effcient. Meanwhile, it greatly enhances the accuracy
and robustness of D-LDA.
5.1.2 Direct Linear Discriminant Analysis
5.1.2.1 The Algorithm of D-LDA
Let S
B
and S
W
denote the between- and the within-class scatter matrices respectively.
The calculation procedure of D-LDA is as follows:
Step 1. Perform eigenvalue decomposition on the between-class scatter matrix
S
B
Let 1
( ,..., )
d
diag =
be the eigenvalue matrix of S
B
in decreasing order and V
= 1
[ ,..., ]
d
v v
be the corresponding eigenvector matrix. It follows that
T
B
V S V = . (5.1)
Let r be the rank of the matrix S
B
. Let
1
[ ,..., ]
r
Y = v v and D
b
=
1
( ,..., )
r
diag
,
and we have
T
B b
Y S Y D =
. (5.2)
Step2. Map each sample vector x to get its intermediate representation
T
Z x using
the projection matrix
1/ 2
b
Z YD
-
=
Step3. Perform eigenvalue decomposition on the within-class scatter matrix of the
projected samples, S
W
which is given by
is prohibited.
T
W W
S Z S Z =
. (5.3)
Let
1
( ,..., )
w r
D diag = be the eigenvalue matrix of
W
S
in ascending order and

U =
1
[ ,..., ]
r
u u
be the corresponding eigenvector matrix. It follows that
T
W w
U S U D =
. (5.4)
Step4. Calculate the discriminant matrix W and map each sample x to
T
W x
The discriminant matrix of D-LDA is given by
1/ 2 1/ 2
b w
W YD UD
- -
=
. (5.5)
5.1.2.2 The Mechanism of D-LDA
From the calculation procedure for discriminant matrix of D-LDA, we fnd that
D-LDA frst maps samples into the range of the between-class scatter matrix using
the matrix Y, and then transforms these projections using a series of regulating
matrices
1/ 2
b
D
-
, U, and
1/ 2
w
D
-
.
The matrix Y itself can be used as a discriminant matrix for feature extraction.
Since it is derived from the Karhunen-Love decomposition of the between-class
scatter matrix, the feature extraction method based on the matrix Y is abbrevi-
ated as KLB. The discriminant matrix of D-LDA W is the multiplication of Y and
1/ 2 1/ 2
b w
D UD
- -
. Thus, D-LDA is in fact a modifed version of KLB.
Now, we study the impact of these regulating matrices on the feature representa-
tion capability of D-LDA.
By expanding the matrix
1/ 2
b
Z YD
-
= we have
1
1 1
1
1
0
1 1
[ ,..., ] [ , , ]
1
0
r r
r
r
Z
(
(
(
= =
(
(
(

v v v v
.
. . . .
.
. (5.6)
This implies that column vectors of Z are weighted column vectors of Y. The
eigenvalue
i
refects the separability of training samples when they are projected
onto the discriminant vector v
i
. The larger the eigenvalue
i
, the better the discrimi-
nant vector v
i
. According to previous assumption,
1
... 0
r
> , the discriminant
vector v
i
is more important than the discriminant vector v
j
, if i > j. Since weight
coeffcients
is prohibited.
1
i
,
1,..., i r =
,
are in inverse proportion to
i
, they deemphasize important discriminant vectors of
KLB. It results in a degraded version of KLB. Thus, the regulating matrix
1/ 2
w
D
-
can-
not enhance the discriminant capability of the matrix Y and should be discarded.
In D-LDA, the matrix Z is regulated by matrices U and
1/ 2
w
D
-
such that its dis-
criminant capability changes greatly.
The expansion of the matrix
1/ 2
w
D
-
is
1
1 1
1
1
0
1 1
[ ,..., ] [ , , ]
1
0
r r
r
r
(
(
(
=
(
(
(

u u u u
.
. . . .
.
. (5.7)
The role of the matrix
1/ 2
w
D
-
is in weighting column vectors of the matrix U.
The eigenvalue
i
also refects the separability of training samples when they are
projected onto the projection vector u
i
. The smaller the eigenvalue
i
, the better the
projection vector u
i
. The projection vector u
i
is more important than the projection
vector u
j
, if
i j
<
. Thus, weight coeffcients

1
i
,
1,..., i r =
,
will emphasize important projection vectors and depress unimportant projection
vectors in some ideal cases.
In other cases, however, when an eigenvalue is either very small or large, the
corresponding weight coeffcient

1
might be either too large or small, and thus over-emphasize projection vectors with
tiny eigenvalues or over-depress projection vectors with huge eigenvalues. In sum,
this is why D-LDA performed well in some cases and failed in others.
is prohibited.
5.1.3.1 The Algorithm of PD-LDA
The following describes in detail the calculation procedure of PD-LDA:
Step1. Perform eigenvalue decomposition on the between-class scatter matrix S
B

to obtain the discriminant matrix of KLB, Y
Step2. Map each sample vector x to obtain its intermediate representation Y
T
x
Step 3. Perform eigenvalue decomposition on the within-class scatter matrix of
projected samples,
W
S
which is given by
T
W W
S Y S Y = . (5.8)
Let 1
( ,..., )
r
diag =
be the eigenvalue matrix of
W
S
in ascending order and

U =
1
[ ,..., ]
r
u u be the corresponding eigenvector matrix. It follows that
T
W
U S U = . (5.9)
Step4. Choose a parameter value and calculate the regulating matrix ( ) using
the following formula
1
( ) ((1 ) ,..., (1 ) )
r
diag = + + . (5.10)
Here, the parameter is used to emphasize projection vectors with small eigen-
values and to depress projection vectors with large ones.
It should be pointed out that the constant one in weight coeffcients
(1 )
i
+
, i
= l,...,r, plays a key role in preventing the over-emphasis of projection vectors with
tiny eigenvalues.
Step5. Calculate the discriminant matrix W and map each sample x to W
T
x
The discriminant matrix of PD-LDA is given by
( ) W YU = . (5.11)
is prohibited.
To alleviate the computational burden in high-dimensional input space, the ma-
trix Y can be calculated using the singular value decomposition theorem as shown
in Yu and Yang (2001).
5.1.3.2 Discussions on the Parameter
Obviously, how to tune the parameter is a crucial problem for PD-LDA. As dis-
cussed in Section 5.1.2.2, the weight coeffcient
1
(1 ) + for u
1
should be larger than
the weight coeffcient (1 )
r
+ for
r
u . It follows that 0 .
When 0 = , the discriminant matrix of PD-LDA is W = YU. Since U is an
orthogonal matrix, the discriminant matrix YU is equivalent to Y in terms of the
recognition rate of a nearest neighbor classifer using Euclidean distance. Therefore,
PD-LDA degenerates to KLB if 0 = .
As weighting coeffcients, (1 )
i
+ , 1,..., i r = , what important is not their abso-
lute but relative values. The ratio of the maximum weight coeffcient
(1 )
i
+
to
the minimum weight coeffcient (1 )
i
+ is crucial for the discriminant capability
of PD-LDA. Let
1
(1 ) /(1 )
r
= + +
. (5.12)
Then, the parameter is given by
1
log( ) / log[(1 ) /(1 )]
r
= + +
. (5.13)
It is obvious that 1 . If 1 = , the regulating matrix ( ) is an identity ma-
trix. In this case all eigenvectors of
W
S are equally weighted. With the increase of
, the representation capability of PD-LDA is gradually enhanced due to properly
emphasizing eigen-vectors of
W
S corresponding to small eigenvalues. When is
large enough, with the increase of , the representation capability of PD-LDA is
gradually weakened due to over-emphasizing. As a result, the curve of recognition
rate of PD-LDA is similar to a parabola.
Now we try to experientially estimate the optimal value of .
In the following experiment, we use the ORL face image database (please refer
to Section 3.2.3). Five randomly selected images of each person are used for train-
ing and the remaining fve for testing. Thus the total amount of training samples
and testing samples are both 200. There is no overlapping between the training
set and the testing set. Figure 5.1 displays the curve of average recognition rate of
PD- LDA with varying over ten runs. Here, the nearest neighbor classifer with
Euclidean distance is used in the experiment.
is prohibited.
From Figure 5.1 we fnd that the curve of the average recognition rate is nearly
a parabola as we expect, and it achieves the maximum around 5. Apparently, the
optimal value of (and the corresponding optimal value of ) might be database-
dependent. It even depends on the partition of a dataset. The value 5 is only an
rough approximation of the optimal value of .
For simplicity, we let = 5 and the parameter be calculated using the formula
(5.13) in all the experiments in this book. An astounding fact is that although the
parameter has not been fnely tuned, the recognition rates of PD- LDA are sig-
nifcantly higher than those of D-LDA, KLB, Eigenface, and Fisherfaces.
The proposed PD-LDA feature extraction method is used for face recognition and
tested on two benchmark datasets, i.e., the AR and FERET face image databases.
To evaluate PD-LDA properly, we also include experimental results for D-LDA,
KLB, and two benchmark facial feature extraction methods, Eigenface (Turk &
Pentland, 1991) and Fisherface (Belhumeur, Hespanha, & Kriengman, 1997). In all
experiments, we use the nearest neighbor classifer with Euclidean distance.
5.1.4.1 Experimental Results on the AR Database
The 120 individuals each having 26 face images taken in two sessions of the AR
face database (please refer to Section 3.4.4) were selected and used in our experi-
ment. Only the full facial images were considered here. The face portion of the image
1 2 3 4 5 6 7 8 9 10
0. 94
0. 945
0. 95
0. 955
0. 96
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
Figure 5.1. Average recognition rate of PD-LDA vs. the ratio of the maximum weight
coeffcient to the minimum weight coeffcient
is prohibited.
was manually cropped and normalized to 5040 pixels as in Yang, Zhang, Frangi
and Yang (2004).
In this experiment, images from the frst session are used for training, and im-
ages from the second session are used for testing. Thus, the total number of train-
ing samples was 840. Since the two sessions were separated by an interval of two
weeks, the aim of this experiment was to compare the performance of PD-LDA
with other facial feature extraction methods under the conditions where there are
changes over time.
Figure 5.2 demonstrates how recognition rates of D-LDA, KLB, Eigenface,
Fisherface and PD-LDA vary with the number of extracted features on the AR
face image database. Here, the Fisherface is the LDA performed on the subspace
spanned by the frst 120 prinicipal components.From Figure 5.2 we fnd that PD-
LDA signifcantly outperforms D-LDA and Fisherface, as well as or a little better
than KLB and Eigenface.
5.1.4.2 Experimental Results on the FERET Database
The same subset of the FERET database as shown in Section 3.4.4 are used. All
images are also resized and pre-processed using the same approch as shown in
Section 3.4.4. The experiment consists of fve tests and each test consists of seven
Figure 5.2. Recognition rates of various methods vs. the number of extracted features
on the AR face image database
19 29 39 49 59 69 79 89 99 109 119
0.45
0.5
0.55
0.6
0.65
0.7
A
v
e
r
a
g
e
r
e
c
o
g
n
i
t
i
o
n
r
a
t
e
PD-LDA
D-LDA
KLB
Eigenface
Fisherface
is prohibited.
runs. In each run of the ith test, i images of each individual are used for training
and the remaining (6 - i) images for testing. Images of each person numbered 1
to (i + 1), 2 to (i + 2), , 7 to i are used as training samples in the frst, second,...,
senventh run respectively.
Table 5.1 shows average recognition rates and standard deviations of D-LDA,
KLB, Eigenface, Fisherface, and PD-LDA in the test on the FERET face image
database. That is, in each of seven runs, four images of each individual are used for
training and the remaining three images for testing. Here the number of extracted
features for each method assumes its maximum value. From Table 5.1, we see
that the PD-LDA is the most accurate method among the fve tested facial feature
extraction methods.
Figure 5.3. Average recognition rates of various methods vs. the number of training
samples per individual on the FERET face image database
Table 5.1. Average recognition rates and standard deviations of different methods
in the test on the FERET dataset
Method D-LDA KLB Eigenface Fisherface PD-LDA
No. of features 199 199 799 199 199
Mean 0.4762 0.5910 0.5619 0.5219 0.6445
Std 0.1305 0.0640 0.0070 0.1160 0.0890
2 3 4 5 6
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Number of training samples per individual
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
PD-LDA
D-LDA
KLB
Eigenface
Fisherface
is prohibited.
Figure 5.3 shows how average recognition rates of D-LDA, KLB, Eigenface,
Fisherface and PD-LDA vary with the number of training samples per individual on
the FERET face image database. In each test, the Fisherface is the LDA performed
on the subspace spanned by the frst 200 prinicipal components. From Figure 5.3
we fnd that PD-LDA outperforms the other four facial feature extraction methods
no matter how many training samples per individual are used.
5.2 Weighted Nullspace Linear Discriminant Analysis
5.2.1 Related Works
Nullspace LDA (N-LDA) (Chen, Liao, Ko, Lin, & Yu, 2000) is one of the most
important variations on linear discriminant analysis (LDA) for SSS problems. Since
standard LDA methods are not applicable to SSS problems due to the singular
within-class scatter matrix S
W
. N-LDA avoids this problem by frst compressing the
high-dimensional input space into the nullspace of S
W
, i.e. N(S
W
) and then seeking
a discriminant matrix which maximizes the between-class scatter in the subspace.
Numerous studies have demonstrated that N-LDA is very accurate for facial feature
extraction and outperforms two popular facial feature extraction methods: Eigenface
and Fisherface. Nonetheless, N-LDA has two drawbacks. First, as it uses a singular
value decomposition of a large scale matrix in its calculation procedure, N-LDA
is computationally rather complex. Second, its separability criterion is not directly
related with classifcation accuracy and as a consequence, as in other LDA methods,
it is not guaranteed that N-LDA will fnd the optimal discriminant matrix.
Many researchers have addressed the issue of the computational complexity of
N-LDA. Chen, Liao, Ko, Lin and Yu (2000) initially used a pixel grouping technique
to alleviate the burden of computation. Huang, Liu, Lu and Ma (2002) proposed
a two-stage N-LDA approach, PCA + N-LDA. This approach frst uses principal
component analysis (PCA) to compress a high-dimensional input space into the
range of the total scatter matrix S
T
, i.e. R(S
T
) and then performs N-LDA on this
reduced subspace. The essence of this approach is performing the singular value
decomposition to the between-class scatter matrix in a much smaller subspace
( ) ( )
T W
R S N S rather than in the high-dimensional subspace N(S
W
). Zheng, Zhao
and Zou (2004) later presented a similar strategy but replaced the orthogonality
constraints with conjugate orthogonality constraints. Recently, Zhao, Yuen and
Yang (2005) theoretically proved that the discriminant matrix of PCA + N-LDA is
equivalent to that of N-LDA.
In the past several years, much work has been done to improve the relationship
between a discriminant criterion and its classifcation accuracy. One of the most
widely-applied approaches has been to introduce a weighting strategy to modify the
is prohibited.
between-class scatter matrix S
B
. Weighting strategies have been used by Lotlikar
and Kothari (2000), Lu, Plataniotis and Venetsanopoulos (2003a), Liang, Gong, Pan
and Li (2005), and Qin, Suganthan and Loog (2006). All of these studies used the
weighted between-class scatter matrix
1
1 1
( )( )( )
c c
T
B i j ij ij i j i j
i j i
S PP w d
-
= = +
= - -
m m m m , (5.14)
to substitute for the between-class scatter matrix
1
1 1
( )( )
c c
T
B i j i j i j
i j i
S PP
-
= = +
= - -
m m m m . (5.15)
Here, P
i
is the class priori probability of the ith class,
( )
ij
w
the monotonically
decreasing function,
1
( ) ( )
T
ij i j W i j
d S
-
= - - m m m m the Mahalanobis distance be-
tween two classes, and m
i
the mean sample of the ith class. The principal reason
for this substitution is that discriminant vectors based on S
B
attempt to preserve
the distances between already well-separated classes while neglecting the small
distances of adjacent class pairs.
In this section, we present a novel facial feature extraction method, weighted
nullspace LDA (WN-LDA) which successfully overcomes the two limitations of
N-LDA. As in PCA + N-LDA (Huang, Liu, Lu, & Ma, 2002; Zheng, Zhao, &
Zou, 2004; Zhao, Yuen, & Yang, 2005), WN-LDA also incorporates a PCA stage
into its computation procedure. However, instead of weighting the between-class
scatter matrix, we enhance the discriminative capability of N-LDA by directly
weighting its discriminant vectors. More specifcally, unlike most LDA techniques
which assume discriminant vectors to be isotropic, the presented method assumes
that discriminant vectors are anisotropic and that important discriminant vectors
should be endowed with larger weights. The proposed method is direct and simple
yet very accurate and effcient. Extensive experimental studies conducted on the
benchmark face image database, FERET demonstrate that WN-LDA is more accurate
than most popular facial feature extraction methods such as Eigenface, Fisherface,
N-LDA, and direct LDA (D-LDA).
is prohibited.
5.2.2.1 The Idea behind WN-LDA
While the concept of feature weight has widely been investigated for feature selec-
tion, for example in text representation (Sebastiani, 2002; Gmez, 2003), it is a new
idea in feature extraction.
The initial motivation of weighted nullspace linear discriminant analysis (WN-
LDA) is demonstrated as follows. Let S
B
and S
W
be respectively the between- and
Algorithm 5.1. A general WN-LDA algorithm
Input:A group of labeled training samples x
1
,...x
N
R
d
, the number of extracted fea-
tures r, and the weight function f()
Output:WN-LDA discriminant matrix V R
dr
1. Compute the between-class scatter matrix S
B
, the within-class scatter matrix S
W
, and
the total scatter matrix S
T
.
2. Calculate the matrix P R
dt
of all eigenvectors of S
T
corresponding to nonzero
eigenvalues using the singular value decomposition theorem. Here t is the rank of
S
T
.
3. Work out the between- and within-class scatter matrices in the range of S
T
using the
formula
T
B B
S P S P
,
and
T
W W
S P S P
.
4. Seek the matrix Q R
ts
of all eigenvectors of
W
S
corresponding to zero eigenvalues.

Here s is the dimension of the subspace N(
W
S
).
5. Figure out the between-class scatter matrix in the nullspace of
W
S
using the formula

( ) ( )
T T
B B B
S Q S Q PQ S PQ =

.
6. Derive the eigenvector- and eigenvalue-matrices U = [u
1
,...,u
s
] and
1
( ,..., )
s
diag = of
the matrix
B
S
.
7. Compute the WN-LDA discriminant matrix V using the formula
1 1
[ ,..., ] ( ( ),..., ( ))
r r
V PQ diag f f u u
.
is prohibited.
the within-class scatter matrices,
1
,...,
s
q q an orthonormal basis for the nullspace of
S
W
,
1
[ ,..., ]
s
Q = q q , and
T
B B
S Q S Q =
. Let
1
,...,
s
u u be eigenvectors of the matrix
B
S

corresponding to eigenvalues
1
,...,
s
in descending order, the frst r ( ( )
B
r rank S )
discriminant vectors of N-LDA are
1
,...,
r
u u . The eigenvalue
i
refects the separa-
bility of training samples when they are projected onto the discriminant vector u
i
.
The larger the eigenvalue
i
, the better the discriminant vector u
i
. Maintaining the
previous assumption,
1
...
r
, the discriminant vector u
i
is more important than
the discriminant vector u
j
, if i < j. It is easy to see that the discriminative capability
of a discriminant matrix will be enhanced if important discriminant vectors are
properly emphasized. Thus, WN-LDA tries to enhance the discriminative capability
of N-LDA by multiplying its discriminant matrix
1 2
[ , ,..., ]
r
U = u u u with a diagonal
matrix
1
( ) ( ( ),..., ( ))
r
f diag f f = . Here ( ) 0 f is a monotonically increasing
function. We call the matrix ( ) f the weight matrix,
1
( ),..., ( )
r
f f the feature
weight coeffcients.
5.2.2.2 The WN-LDA Algorithm
The following is a general algorithm for WN-LDA which marries the ideas of PCA
+ N-LDA and feature weight.
5.2.2.3 Choosing the Weight Function
Now, the problem is how to properly choose the weight function ( ) f so as to
improve the discriminative capability of the N-LDA discriminant matrix. In this
chapter, we are going to investigate a family of simple weight functions with one
parameter, i.e. power functions ( ) f = with 0 > . These functions are mono-
tonically increasing such that important discriminant vectors are endowed with
larger weight coeffcients.
Apparently, absolute values of feature weight coeffcients are not important and
what is really important is their relative scores. Besides the form of a weigh func-
tion, the discriminative capability of WN-LDA crucially depends upon the ratio
of the maximum feature weight coeffcient
1
( ) f
to the minimum feature weight
coeffcient
( )
r
f
,
1
( )
( )
r
f
f
=
.
It is obvious that the ratio is not less than 1. If 1 = , the weight matrix
( ) f

is equivalent to an identity matrix such that all discriminant vectors of N-LDA are
equally weighted. As a consequence, WN-LDA degenerates into N-LDA. Generally
speaking, the optimal value of is database dependent. It depends on the partition
is prohibited.
Figure 5.4. ARR curves of WN-LDA over the ratio with different numbers of
classes
1 1.5 2 2.5 3 3.5
0.65
0.7
0.75
0.8
0.85
0.9
Ratio of the maximum feature weight coefficient to the minimum feature weight coefficient
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
c=40
c=80
c=120
c=160
c=200
of a database, the number of extracted features, and other database-related features.
The optimal value of the ratio can be estimated experimentally. Once the ratio
is fxed, the parameter can be computed using the formula
1
log( ) / log( / )
r
=
. (5.16)
The proposed facial feature extraction method was evaluated on the benchmark
face image database, FERET. WN-LDA was evaluated on a series of subsets of
the FERET database. The biggest subset includes 1,400 images of 200 individuals
with seven images of each individual.
In all experiments we used the nearest neighbor (NN) classifer with Euclidean
distance.
Before comparing the accuracy of WN-LDA with that of state-of-the-art facial
feature extraction methods, we frst investigate the impact of , the ratio of the
maximum feature weight coeffcient to the minimum feature weight coeffcient,
on the average recognition rate (ARR) of WN-LDA.
5.2.3.1 Impact of the Parameter with Varying Number of Classes
The experiment consists of fve tests of ten runs each. In the ith test, three images
of each of 40i individual are randomly chosen for training, while the remaining
is prohibited.
four images are used for testing. Thus, the training sample set size is 120i and the
testing sample set size is 160i, i = 1,,5. In this way, we ran the system ten times
and obtained ten different training and testing sample sets for each test. In each
run, the ratio in turn assumes 1, 1.5, 2, 2.5, 3, and 3.5. In addition, in each test,
the number of extracted features assumes the maximum value, i.e. the rank of
between-class scatter matrix.
Figure 5.4 depicts the ARR curves of WN-LDA over the ratio when the number
of classes c is respectively equal to 40, 80, 120, 160, and 200. It can be seen that if
all discriminant features are extracted, the larger the number of classes, the more
sensitive the accuracy of WN-LDA to the ratio .
5.2.3.2 Impact of the Parameter with Varying Number of Extracted
Features
Seven images of each of 200 individuals were used in the experiment. Three im-
ages of each individual were randomly chosen for training, while the remaining
four images were used for testing. Thus, the training sample set size was 600 and
the testing sample set size was 800. In this way, we ran the system ten times and
obtained ten different training and testing sample sets.
Figure 5.5 displays the ARR curves of WN-LDA over the number of extracted
features when the ratio assumes different values. Identifers WN-LDA2, , WN-
LDA6 in Figure 5.5 stand for WN-LDA with the ratio being equal to 2, , 6.
9 14 19 24 29 24 39 49 79 109 139 169 199
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
N-LDA
WN-LDA2
WN-LDA3
WN-LDA4
WN-LDA5
WN-LDA6
Figure 5.5. The ARR curves of WN-LDA over the number of extracted features when
the ratio assumes different values
is prohibited.
From Figure 5.5 we fnd that, unlike N-LDA whose ARR is heavily dependent
on the number of extracted features, the accuracy of WN-LDA is quite robust.
Furthermore, with the increase of the ratio , the ARR curve of WN-LDA is ap-
proaching a monotonically increasing curve. One probable reason for this is that
the WN-LDA discriminant vectors corresponding to minor eigenvalues might be
incorrect due to roundoff error and poor conditioning of the between-class scatter
matrix, and smaller feature weight coeffcients properly depress the side effects of
these discriminant vectors.
5.2.3.3 Comparison of WN-LDA with Other Facial Feature
Extraction Methods
The experimental design is the same as in Section 5.2.3.2. Figure 5.6 shows the
ARR curves of WN-LDA ( 2 = ), N-LDA, D-LDA, Eigenface, and Fisherface.
It can be seen that WN-LDA is more accurate than the other four facial feature
extraction methods.
5.2.3.4 Experimental Results over Varying Number of Training
Samples per Individual
According to Lu, Plataniotis and Venetsanopoulos (2005) the accuracy of D-LDA
is heavily dependent on the number of training samples per individual. To fairly
Figure 5.6. Average recognition rate of various methods vs. the number of extracted
features
19 49 79 109 139 169 199
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
WN-LDA
N-LDA
D-LDA
Eigenface
Fisherface
is prohibited.
compare WN-LDA with D-LDA and other methods, in this subsection we evaluate
the accuracy of WN-LDA ( 2 = ) with that of N-LDA, D-LDA, Eigenface, and Fish-
erface when the number of training samples per individual varies from 2 to 6.
The experiment consists of fve tests of ten runs each. In each run of the ith
test, (i + 1) images of each individual were randomly selected for training and the
remaining (6 - i) images for testing. The number of extracted features for WN-LDA,
N-LDA, D-LDA, Fisherface, and Eigenface in the ith test were respectively 15i,
15i,15i,15i,and200 199 i + .Figure 5.7 displays the ARR curves of various feature
extraction methods over varying number of training samples per individual. It can
be seen that WN-LDA is once again the most accurate method among the tested
facial feature extraction approaches.
5.3 WEight ED LiNEAR DiSCRiMiNANt ANALySiS iN th E
RANgE OF Withi N-CLASS SCAtt ER MAt Rix
Conventional facial feature extraction methods including Fisherface, N-LDA, and
D-LDA perform fairly well in small uniform face image databases such as Yale,
UMIST, ORL, and AR. But, if applied to large noisy datasets such as FERET, they
suffer from two particular problems. First, they are not very accurate. Second,
while their accuracy is heavily dependent on the number of extracted features, the
optimal number of extracted features is hard to determine.
Figure 5.7. Average recognition rates of various feature extraction methods vs. the
numbers of training samples per individual
2 3 4 5 6
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e WN-LDA
N-LDA
D-LDA
Eigenface
Fisherface
is prohibited.
To overcome these two drawbacks we propose a novel facial feature extraction
method, weighted linear discriminant analysis in the range of the within-class
scatter (WLDA-RWS) matrix in this section. We began by frst extensively testing
simple LDA in the range of the within-class scatter matrix (LDA-RWS) on the
FERET benchmark face image database, demonstrating that LDA-RWS generally
outperforms N-LDA, D-LDA, and Fisherface. Further tests properly weighting the
discriminant vectors of LDA-RWS with feature weight coeffcients further improved
both its discriminative capability and its stability.
This research also made some discoveries which appear to overturn two very
common beliefs about discriminant feature extraction. First, contradicting the pre-
vailing view that, given a small sample size, optimal Fisher discriminant vectors
should be derived from the nullspace of the within-class scatter matrix, we found
that, when applied to a collection of large scale noisy face images, a set of Fisher
discriminant vectors derived from the range of the within-class scatter matrix was
more accurate than the set of discriminant vectors of N-LDA. This implies that
larger Fisher criterion function values do not guarantee greater accuracy. Second,
we discovered that discriminant matrices with the same discriminant criterion
value, i.e. they are all the optimal solutions to the model (4.1) are unnecessarily of
the same discriminative capability. It is well known that if
* d r
W R

is the Fisher
discriminant matrix and
r r
R

an arbitrary nonsingular matrix, then the matrix
*
W is also an optimal solution to the model (4.1). In the past, researchers com-
monly believed that the matrix had no effect on the discriminative capability of
W
*
and that thus the matrix
*
W was equivalent to the matrix W
*
. This implied that
discriminant vectors (directions) are isotropic. However, our experimental results
have shown that proper selection of a weight matrix
1
( ,..., )
r
diag f f =
offers the
potential to greatly improve the discriminative capability in terms of recognition
accuracy. Here,
1
,...,
r
f f are feature weight coeffcients.
5.3.1.1 The Ideas behind LDA-RWS
Let N(S
W
) and R(S
W
) be both the nullspace and the range of the within-class scatter
matrix. According to linear algebraic theory, the input space R
d
can be decomposed
as the direct sum of these two subspaces, i.e. ( ) ( )
d
W W
R N S R S = . To avoid the
singularity problem, LDA-RWS converts the Fisher discriminant criterion into
( ) ( )
max ( )
W W
T
B
T
W R S R S
W
W S W
J W
W S W

=
. (5.17)
is prohibited.
There always is an optimal solution to the optimization model (5.17). In fact,
if P is the matrix comprised of unit eigenvectors of the matrix S
W
corresponding
to nonzero eigenvalues, and Q
r
is the matrix composed of unit eigenvectors of the
matrix
1
( ) ( )
T T
W B
P S P P S P
-
corresponding to the r largest eigenvalues, then PQ
r
is an
optimal solution to the model (2). Furthermore, here PQ
r
is used as the discriminant
matrix of LDA-RWS. Let
1
,...,
r
be the r largest eigenvalues of
1
( ) ( )
T T
W B
P S P P S P
-
. We have
1
( )
r
r i
i
J PQ
=
=
. Let W be the discriminant matrix of N-LDA. It follows

that ( ) ( )
W W
W N S N S . Thus, we have ( ) ( )
r
J W J PQ = > . This is the primary
reason why many researchers believe that the optimal Fisher discriminant vectors
should be derived from N(S
W
)

rather than from R(S
W
)

. A detailed discussion of the
relevant experimental results can be found in the following subsection.
5.3.1.2 The Concept of Feature Weight
The concept of using feature weights for feature selection, for example in areas
such as text representation (Sebastiani, 2002; Gmez, 2003), has been widely in-
vestigated. It appears, however, to be new in feature extraction. One reason for this
omission may be that many researchers in the feld of dimensionality reduction or
feature extraction take it for granted that two transformation matrices with the same
discriminant criterion value should have the same discriminative capability.
Let W be one discriminant matrix and
1
,...,
r
f f to be a set of feature weight coef-
fcients. It is obvious that following equations satisfy
1 1
1
1 1
( ,..., ) ( ,..., )
( ( ,..., ))
( ,..., ) ( ,..., )
( )
T
r B r
r
T
r W r
T
B
T
W
diag f f W S Wdiag f f
J Wdiag f f
diag f f W S Wdiag f f
W S W
J W
W S Wdiag
= =
=
(5.18)
This would suggest that feature weights should have no impact on the discrimina-
tive capability of a discriminant matrix. Yet Song, Liu and Yang (2005) have shown
that although orthogonalized Fisher discriminant matrix and Fisher discriminant
matrix have the same discriminant criterion value, in terms of recognition accuracy
there is a signifcant difference in their discriminative capabilities.
Further, it is well known that with the increase of number of classes and noise
in the data, discriminant vectors with minor eigenvalues become less accurate.
As a consequence, average recognition accuracies of various feature extraction
methods do not increase with the number of extracted features when associated
eigenvalues of newly added discriminant vectors are small enough. This implies
is prohibited.
that discriminant vectors are anisotropic at least under circumstances with large-
scale noisy data.
Let P be the matrix comprised of unit eigenvectors of the matrix S
W
corresponding
to nonzero eigenvalues, and
1
,...,
w
q q eigenvectors of the matrix
1
( ) ( )
T T
W B
P S P P S P
-

corresponding to eigenvalues
1
,...,
w
in descending order, the frst r ( ( )
B
r rank S )
discriminant vectors of LDA-RWS are
1
,...,
s
P P q q . The eigenvalue
i
refects the
separability of training samples when they are projected onto the discriminant
vector
i
Pq
. The larger the eigenvalue
i
, the better the discriminant vector
i
Pq
.
According to the previous assumption,
1
...
r
, the discriminant vector
i
Pq is
more important than the discriminant vector
i
Pq , if i j < . Clearly, the discrimina-
tive capability of a discriminant matrix will be enhanced if important discriminant
vectors are properly emphasized.
5.3.1.3 The WLDA-RWS Algorithm
WLDA-RWS tries to enhance the discriminative ability of LDA-RWS by multiplying
its discriminant matrix
r
PQ with a diagonal matrix
1
( ) ( ( ),..., ( ))
r
f diag f f = .
Algorithm 5.2. A general WLDA-RWS algorithm
Input:A group of labeled training samples x
1
,...,x
N
R
d
, the number of extracted features
r, and the weight function f()
Output:WLDA-RWS discriminant matrix V R
d r
Step1. Compute the between-class scatter matrix S
B
and the within-class scatter matrix
S
W
.
Step2. Calculate the matrix P R
d w
of all eigenvectors of S
w
corresponding to nonzero
eigenvalues using the singular value decomposition theorem. Here w is the rank of S
w
.
Step3. Work out the between- and within-class scatter matrices in the range of S
w
using
the formulae
T
B B
S P S P
,
and
T
W W
S P S P
.
Step4. Seek the eigenvector- and eigenvalue-matrices Q = [q
1
,...,q
w
] and
1
( ,..., )
w
diag = of the matrix
1
W B
S S
-

.
Step5. Compute the WN-LDA discriminant matrix V using the formula
1 1
[ ,..., ] ( ( ),..., ( ))
r r
V P diag f f q q
.
is prohibited.
Step1. Calculate the slope a
p
and the intercept b
p
of the previous line segment using the
formulae
1
( 1) /( )
p r
a - - ,
and
1
p p r
b a -
.
Step2. Compute the x- and y-coordinates of the right end, and the x-coordinate of the left
end of the current line segment using the formulae
1
_ right x ,
_ right y ,
and
1
_ ( 1)
r
r
left x
-
+ - .
Step3. Derive the slope a
c
and the intercept b
c
of the current line segment using the
formulae
_ _ left y a left x b + ,
_ _ ( _ _ ) left y left y right y left y + - ,
_ _
_ _
c
right y left y
a
right x left x
-
-
,
and
_ _
c c
b right y a right x - .
Step4.Work out ( ) f for _ _ left x right x < using the formula
( )
c c
f a b + .
Step5. Update intermediate variables using the formulae
( _ 1) /( _ )
p r
a left y left x - - ,
Algorithm 5.3. Algorithm for the pairwise linear weight function
Input:A set of positive eigenvalues
1
,...,
r
in descending order, the ratio of the maxi-
mum feature weight coeffcient to the minimum feature weight coeffcient , the number
of line segments , and the lift rate of the pairwise linear curve
Output:A set of feature weight coeffcients
1
( ),..., ( )
r
f f
is prohibited.
1
p p r
b a - ,
_ _ right x left x ,
_ _ right y left y ,
and
1
_ _
r
left x left x
-
-
.
Step6. Repeat Step 3 to Step 5 until all feature weight coeffcients have been obtained.
Here ( ) 0 f is a monotonically increasing function. We call the matrix ( ) f the
weight matrix, and
1
( ),..., ( )
r
f f the feature weight coeffcients.
The following is a general algorithm for WLDA-RWS which marries the ideas
of LDA-RWS and feature weight.
5.3.1.4 Pairwise Linear Weight Functions
Now, the problem is how to properly choose the weight function ( ) f such that the
discriminative capability of the LDA-RWS discriminant matrix can be enhanced.
In this paragraph , we are going to investigate a family of pairwise linear weight
functions having three parameters , , and .
The frst parameter is the ratio of the maximum feature weight coeffcient
1
( ) f
to the minimum feature weight coeffcient
( )
r
f
, i.e.
1
( )
( )
r
f
f
=
.
It is obvious that the ratio is a positive number not less than 1. If = 1, the weight
matrix ( ) f is equivalent to an identity matrix such that all discriminant vectors
of LDA-RWS are equally weighted. As a consequence, WLDA-RWSM degener-
ates into LDA-RWS. The second parameter is the number of line segments in the
pairwise linear curve. As a result, is a positive integer. If = 1, the pairwise linear
curve degenerates into a line. The third parameter is the lift rate of the pairwise
linear curve which determines the lift distance of the left end of the current line
segment relative to the corresponding point of the previous line segment. is a
is prohibited.
Figure 5.8. Geometric explanation of the parameters , , and
d
) 2 (
1
-
-
+
r
r
D
) 1 (
1
-
-
+
r
r
1
D
d
1
2
3
4
( , , )=(2,2,0.5)
( , , )=(3,2,0.5)
( , , )=(4,2,0.5)
number between zero and one. If = 1, the pairwise linear curve degenerates into
a line whereas if = 1, the pairwise linear curve degenerates into a horizontal line.
As a consequence, WLDA-RWS degenerates into LDA-RWS. Figure 5.8 gives the
geometric explanation of the parameters , , and .
The detailed algorithm for computing feature weight coeffcients of the pairwise
linear weight function is as follows.
Figures. 5.9-5.11 plot some sample pairwise linear curves. All these pairwise
linear weight functions are monotonically increasing. As a consequence, important
discriminant vectors can be endowed with larger weight coeffcients.
Figure 5.9. Sample pairwise linear curves with varying values of
is prohibited.
r 1
1
1.2
1.4
1.6
1.8
2
( , , ) = (2,2,0.5)
( , , ) = (2,3,0.5)
( ) = (2 4 0 5)
r 1
1
1.2
1.4
1.6
1.8
2
( , , ) = (2,3,0.50)
( , , ) = (2,3,0.67)
( , , ) = (2,3,0.75)
The proposed facial feature extraction method was evaluated on the benchmark face
image database, FERET. WLDA-RWS was evaluated on a subset of the FERET
database (please refer to Section 3.4.4). The subset includes 1,400 images of 200
individuals with seven images of each individual.
is prohibited.
5.3.2.1 Comparison of WLDA-RWS with Other Facial Feature
Extraction Methods
Three images of each individual were randomly chosen for training and the remain-
ing four images were used for testing. Thus, the training sample set size is 600 and
the testing sample set size is 800. We ran the system ten times and obtained ten
different training and testing sample sets.
Figure 5.12. Average recognition accuracy of various methods vs. the number of
extracted features
Figure 5.13. Average recognition accuracy of various feature extraction methods
vs. the numbers of training samples per individual
9 19 29 39 49 79 109 139 169 199
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

a
c
c
u
r
a
c
y
N-LDA
D-LDA
Fisherface
LDA-RWS
WLDA-RWS
2 3 4 5 6
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

a
c
c
u
r
a
c
y
N-LDA
D-LDA
Fisherface
LDA-RWS
WLDA-RWS
is prohibited.
Figure 5.12 shows the average accuracy curves of N-LDA, D-LDA, Fisherface,
LDA-RWS, and WLDA-RWS having parameters ( , , ) (2, 7, 0.9) = . It can be
seen that WLDA-RWS is generally more accurate than the other four facial feature
extraction methods.
As mentioned in Section 5.2.3.4, the accuracy of D-LDA varies with the number
of training samples per individual. In this subsection we evaluate the accuracies
of WLDA-RWS, N-LDA, D-LDA, Fisherface, and LDA-RWS under the condition
that the number of training samples per individual ranges between 2 and 6. This
seems to be a fair and reasonable scheme to compare WLDA-RWS with D-LDA
and other methods.
The experiment consists of fve tests of ten runs. In each run of the ith test, (i +
1) images of each individual were randomly selected for training and the remaining
(6 i) images were retained for testing. There are 199 extracted features for each
of N-LDA, D-LDA, Fisherface, LDA- RWS, and WLDA-RWS.
Figure 5.13 shows the average accuracy curves of various feature extraction
methods over varying number of training samples per individual. It can be seen
that among the tested facial feature extraction approaches, WLDA-RWS is again
the most accurate.
5.4 summAr y
In this chapter we propose three kinds of parameterized linear discriminant analysis
methods. The frst is parameterized direct linear discriminant analysis; the second
is weighted nullspace linear discriminant analysis; and the third is weighted linear
discriminant analysis in the range of within-class scatter matrix. The common points
are that they all combine the concept of feature weight with existing LDA methods
for small sample size problems to further improve their discriminant capability.
r eferences
Belhumeur, P. N., Hespanha, J. P., & Kriengman, D. J. (1997). Eigenfaces vs. fsh-
erfaces: recognition using class specifc linear projection. IEEE Trans. on Pattern
Anal. Machine Intell., 19(7), 711-720.
is prohibited.
33(10), 1713-1726.
Gmez, J. M. (2003). Text representation for automatic text categorization. Pro-
ceedings of Eleventh Conference of the European Chapter of the Association for
Computational Linguistics.
Huang, R., Liu, Q. S., Lu, H. Q., & Ma, S. D. (2002). Solving the small sample size
problem of lda. Proceedings of International conference on Pattern Recognition,
(pp. 29-32).
Liang, Y. X., Gong, W. G., Pan, Y. J., & Li, W. H. (2005). Generalizing relevance
weighted lda. Pattern Recognition, 38(11), 2217-2219.
Lotlikar, R., & Kothari, R. (2000). Fractional-step dimensionality reduction. IEEE
Trans. Pattern Anal. Machine Intell., 22(6), 623-627.
Lu, J., Plataniotis, K., & Venetsanopoulos, A. (2003a). Face recognition using lda-
based algorithms. IEEE Trans. on Neural Networks, 14(1), 195-200.
Lu, J., Plataniotis, K., & Venetsanopoulos, A. (2003b). Regularized discriminant
analysis for the small sample size problem in face recognition. Pattern Recognition
Lett., 24(16), 3079-3087.
Lu, J., Plataniotis, K., & Venetsanopoulos, A. (2003c). Face recognition using
kernel direct discriminant analysis algorithms. IEEE Trans. on Neural Networks,
14(1), 117-126.
Lu, J., Plataniotis, K., & Venetsanopoulos, A. (2005). Regularization studies of
linear discriminant analysis in small sample size scenarios with application to face
recognition. Pattern Recognition Letter, 26(2), 181-191.
Lu, J., Plataniotis, K., Venetsanopoulos, A., & Li, S. (2006). Ensemble-based
discriminant learning with boosting for face recognition. IEEE Trans. on Neural
Networks, 17(1), 166- 178.
Qin, A. K., Suganthan, P. N., & Loog, M. (2006). Generalized null space uncor-
related Fisher discriminant analysis for linear dimensionality reduction. Pattern
Recognition, 39(9), 1805-1808.
Sebastiani, F. (2002). Machine learning in automated text categorization. ACM
Computing Surveys, 34(1), 1-47.
Song, F. X., Liu, S. H., & Yang, J. Y. (2005). Orthogonalized Fisher Discriminant.
is prohibited.
Song, F. X., Zhang, D., Wang, J. Z., Liu, H., & Tao, Q. (2007). A parameterized direct
LDA and its application to face recognition. Neurocomputing, 71(1-3), 191-196.
Turk, M., & Pentland, A. (1991). Face recognition using Eigenfaces. Proc. IEEE
Conf. Computer Vision and Pattern Recognition, (pp. 586-591).
Yang, J., Zhang, D., Frangi, A. F., & Yang, J. Y. (2004). Two-dimensional pca: a new
approach to appearance-based face representation and recognition. IEEE Trans. on
Pattern Anal. Machine Intell., 26(1), 131-137.
Yu, H., & Yang, J. (2001). A direct lda algorithm for high-dimensional datawith
Zhao, H. T., Yuen, P. C., & Yang, J. Y. (2005). Optimal subspace analysis for face
recognition. International Journal of Pattern Recognition and Artifcial Intelligence,
19(3), 375- 393.
Zheng, W. M., Zhao, L., & Zou, C. R. (2004). An effcient algorithm to solve the
small sample size problem for lda. Pattern Recognition, 37(5), 1077-1079.
is prohibited.
Chapter VI
Two Novel Facial Feature
Extraction Methods
Abstr Act
In this chapter, we introduce two novel facial feature extraction methods. The frst is
multiple maximum scatter difference (MMSD) which is an extension of a binary linear
discriminant criterion, i.e. maximum scatter difference. The second is discriminant
based on coeffcients of variances (DCV) which can be viewed as a generalization
of N-LDA. At last, we give a summery of the chapter.
6.1 mul t IPle mAxImum scAtter dI fference
The maximum scatter difference (MSD) discriminant criterion (Song, Zhang,
Chen, & Wang, 2007) presented in Section 3.4 is a binary discriminant criterion
for pattern classifcation. Because MSD utilizes the generalized scatter difference
rather than the generalized Rayleigh quotient as a class separability measure, it
avoids the singularity problem when addressing the SSS problems that trouble the
Fisher Discriminant Criterion. Further, experimental studies demonstrated that
MSD classifers based on this discriminant criterion have been quite effective on
face recognition tasks (Song et al., 2007). The drawback of the MSD classifer is
that, as a binary classifer, it cannot be applied directly to multiclass classifcation
problems such as face recognition. This means that multiple recognition tasks
have to be divided into a series of binary classifcation problems using one of three
implementation strategies: one-vs-rest, one-vs-one, or directed-acyclic-graph (Hsu
Two Novel Facial Feature Extraction Methods 107
is prohibited.
& Lin, 2002). Experiments have shown that MSD classifers are not very effective
when using the frst strategy, while using the latter two strategies requires the train-
ing of l(l-1)/2 MSD classifers for a l-class recognition problem. The effciency of
such an approach will greatly be affected by any increase in the number of classes.
Ultimately, then, like all binary classifers, MSD classifers are not suitable for
large-scale pattern recognition problems.
To address the problem, this section generalizes the classifcation-oriented binary
criterion to its multiple counterpartmultiple maximum scatter difference (MMSD)
discriminant criterion for facial feature extraction (Song, Liu, & Yang, 2006;
Song, Zhang, Mei, & Guo, 2007). The MMSD feature extraction method based
on this novel discriminant criterion is a new subspace-based feature extraction
method. Unlike most conventional subspace-based feature extraction methods that
derive their discriminant vectors either in the range of the between-class scatter
matrix or in the null space of the within-class scatter matrix, MMSD computes its
discriminant vectors in both subspaces. MMSD is theoretically elegant and easy
to calculate. Extensive experimental studies conducted on the benchmark data-
base, FERET, show that MMSD outperforms many state-of-the-art facial feature
extraction methods including nullspace LDA (N-LDA) (Chen, Liao, Ko, Lin, &
Yu, 2000), direct LDA (D-LDA) (Yu & Yang, 2001), Eigenface (Turk & Pentland,
1991), Fisherface (Belhumeur, Hespanha, & Krieqman, 1997), and complete LDA
(Yang, Frangi, Yang, Zhang, & Jin, 2005).
6.1.1 Related Works
Almost all facial feature extraction methods have been developed from Fisher linear
discriminant (FLD). FLD seeks an optimal linear transformation
* d r
W R

from
a high-dimensional input space R
d
into a low-dimensional feature space R
r
, by
observing the following Multiple Fisher Discriminant Criterion,
max
d r
T
B
T
W R
W
W S W
W S W
. (6.1)
Here S
B
and S
W
are respectively the between- and within-class scatter matrices.
The transformation matrix W
*
is called the Fisher discriminant matrix and its
column vectors are called Fisher discriminant vectors. It has been proved that the
matrix whose ith column vector is an eigenvector of
1
W B
S S
-
ith largest eigenvalue is an optimal solution for the model (6.1) when S
W
is nons-
ingular. Unfortunately, in SSS problems such as face recognition, the dimension
of an input space d is usually greater than the number of training samples N. As
is prohibited.
a result, the within-class scatter matrix S
W
degenerates into a singular matrix. In
such a case, model (6.1) is ill-posed and therefore FLD cannot be directly applied
to face recognition.
Various facial feature extraction methods have been proposed based on one or
another modifcation of FLD. These methods fall roughly into two groups: com-
putation-oriented and subspace-based. Computation-oriented methods seek an
approximate substitute for the inverse of the within-class scatter matrix when it is
singular. Well known computation-oriented methods include the pseudo inverse
(Tian, Barbero, Gu, & Lee, 1986), singular value disturbation (Hong & Yang, 1991),
and regularization (Zhao, Chellappa, & Phillip, 1999). Because computation-oriented
approaches incur a heavy burden of calculation and are not effective they are no
longer the focus of studies.
Subspace-based approaches compress the feasible region of discriminant vectors
from the total input space into one subspace such that discriminant matrices can
be easily derived in this subspace. Early examples include Fisherface (Belhumeur
et al., 1997), the nullspace LDA (N-LDA) (Chen et al., 2000), and direct LDA
(D-LDA) (Yu & Yang, 2001). Fisherface calculates its discriminant vectors in the
subspace spanned by a set of several major eigenvectors of the total scatter matrix.
N-LDA seeks its discriminant vectors in the null space of the within-class scatter
matrix, N(S
W
) whereas D-LDA tries to fnd its discriminant vectors in the range of
the between-class scatter matrix, R(S
B
). Subspace-based feature extraction methods
have fourished in recent years. Numerous studies have shown that both N(S
W
) and
R(S
B
) contain important discriminant information so feature extraction methods
that derive discriminant vectors in both subspaces should perform better. Based
on this belief, Yang et al. proposed complete LDA (C-LDA) (Yang et al., 2005) and
experimental studies have shown that C-LDA does outperform other subspace-based
facial feature extraction methods.
6.1.2.1 The MMSD Discriminant Criterion
The goal of discriminant criteria for feature extraction is to seek r discriminant
vectors w
1
,...,w
r
such that training samples from a high-dimensional input space
are farthest apart after they are projected on these vectors. In order to eliminate the
infuences of lengths of discriminant vectors and linear dependences between these
vectors we usually require them to be orthonormal. That is, discriminant vectors
should satisfy the constraints , , 1,...,
T
i j ij
i j r = = w w .
Let
1
[ ,..., ]
d r
r
W R

= w w be the discriminant matrix. The projection of a
sample x on discriminant vectors w
1
,...,w
r
is W
T
x. The between- and within-class
is prohibited.
scatter matrices of projected training samples
1
,...,
T T
N
W W x x are
T
B B
S W S W =
and
T
W W
S W S W =
respectively.
The trace of the between-class scatter matrix of projected training samples,
( )
T
B
tr W S W refects the scatter of projected training samples between categories
whereas the trace of the within-class scatter matrix of projected training samples,
( )
T
W
tr W S W refects the scatter of projected training samples within each category.
The larger the ( )
T
B
tr W S W , the more separable the projected data; the smaller the
( )
T
W
tr W S W , the more separable the projected data. Thus, we wish to achieve two
distinct objectives:
max ( )
T
B
tr W S W , (6.2)
min ( )
T
W
tr W S W
, (6.3)
satisfying the orthonormal constraints:
, , 1, 2,...,
T
i j ij
i j r = = w w . (6.4)
The problem of seeking a set of discriminant vectors is then translated into a
problem of solving a multi-objective programming model which is defned by (6.2-
6.4). It is well known that a multi-objective programming model cannot be solved
directly. It has to frst be converted into a single-objective programming model.
There are two main ways to convert a multi-objective programming model into a
single-objective model: The frst is the goal programming approach in which one
of the objectives is optimized while the remaining objectives are converted into
constraints. The second is the combining objective approach in which all objec-
tives are combined into one scalar objective. When using a combining objective
approach there are two major rules: the multiplicative and the additive. By utilizing
the multiplicative rule to combine objectives (6.2) and (6.3), we gain a discriminant
criterion that is similar to the generalized Fisher discriminant criterion (Guo, Li,
Yang, Shu, & Wu, 2003). If using the additive rule to combine the two objectives,
we obtain the multiple maximum scatter difference (MMSD) discriminant criterion.
The single-objective optimization model corresponding to the MMSD discriminant
criterion is as follows:
1 2 1 2
, , 1,2,...,
max {[ , ,..., ] ( )[ , ,..., ]}
T
i j ij
T
r B W r
i j r
tr S c S
= =
-
w w
w w w w w w
. (6.5)
It can further be transformed into
is prohibited.
, , 1,2,...,
1
max ( )
T
i j ij
r
T
i B W i
i j r
i
S c S
= =
=
-
w w
w w. (6.6)
Here
B W
S c S - is a generalized scatter difference matrix and c a nonnegative
parameter.
Obviously, model (6.6) is an extension of the model (3.28).
According to Theorem 4 in (Guo et al., 2003), we can conclude that the ortho-
normal eigenvectors
1
,...,
r
of the matrix
B W
S c S - v corresponding to the r largest
eigenvalues make of an optimal solution of model (6.6). We call these orthonormal
eigenvectors MMSD discriminant vectors. It should be pointed out that a similar
discriminant criteriondifferential scatter discriminant criterion has been inves-
tigated in Fukunaga (1990) and extended to tensor minimax probability machines
(Tao, Li, Hu, Maybank, & Wu, 2005) and general tensor discriminant analysis (Tao,
Li, Maybank, & Wu, 2006) .
6.1.2.2 The MMSD-Based Feature Extraction Algorithm
It is expensive both in time and memory to directly perform eigen decomposition on
the matrix
B W
S c S - when the dimensionality of the input space d is large enough.
It is the key of the MMSD-based feature extraction algorithm how to decompose
the matrix
B W
S c S -
quickly. Fortunately, however, with the following lemma and
theorem we can always compute the eigenvectors of
B W
S c S - by performing ei-
gen-decomposition on a much smaller matrix whose dimension is equal to or less
than ( 1) N - ( 1) N - .
Lemma6.1.Let S
B
and S
W
be respectively the between- and within-class scatter
matrices. Let P be the matrix of all unit eigenvectors of the total scatter matrix S
T

(= S
B
+ S
W
) corresponding to nonzero eigenvalues. Then it follows that
(1)
T
B B
PP S S = ,
(2)
T
W W
PP S S =
.
Proof: Let
1
[ ,..., ]
s
Q = q q be the matrix of all unit eigenvectors of the total scatter
matrix corresponding to zero eigenvalues. It is obvious that V = [P Q] is a unitary
matrix.
Since ( 1, 2,..., )
i
i s = q is an eigenvector of S
T
corresponding to zero eigenvalue, it
follows that
T i
S = q 0. Thus we have
T T
i B i i W i
S S + = q q q q 0. In consideration of the fact
that S
B
and S
W
are both semi-positive matrices, it can be concluded that
T
i B i
S = q q 0
is prohibited.
and
T
i W i
S = q q 0. Using the Theorem 2 in Liu, Cheng and Yang (1992a) it follows
that
T
i B
S = q 0 and
T
i W
S = q 0. As a consequence, we have
1
T
B
T T T T T
B B B B B B
T
s B
S
S VV S PP S QQ S PP S Q PP S
S
| |
|
= = + = + =
|
|
\ .
q
q
.
,
and
1
T
W
T T T T T
W W W W W W
T
s W
S
S VV S PP S QQ S PP S Q PP S
S
| |
|
= = + = + =
|
|
\ .
q
q
. .
Theorem6.1.Let
d t
P R

be the matrix of all unit eigenvectors of the total scatter
matrix S
T
corresponding to nonzero eigenvalues and
1 t
R

to be the eigenvector
of the matrix ( )
T
B W
P S c S P - corresponding to the eigenvalue . Then P is the
eigenvector of the matrix
B W
S c S - corresponding to the eigenvalue .
Proof: Since is the eigenvector of the matrix ( )
T
B W
P S c S P - corresponding to
the eigenvalue , we have
( ) ( )
( )
T T
B W B W
T T
B W
P S c S P PP S c S P P
PP S c PP S P P
- = - =
- =
.
From Lemma 6.1, it follows that ( )
B W
S c S P P - = .
Thus, we complete the proof of the theorem.
According to Theorem 6.1, we can calculate the MMSD discriminant vectors in a
high-dimensional input space in three major steps: First, we calculate the matrix of
Algorithm 6.1. Facial feature extraction algorithm based on MMSD
1. Compute the between-class scatter matrix S
B
, the with-class scatter matrix S
W
, and the
total scatter matrix S
T
.
2. Calculate the matrix P of all unit eigenvectors of S
T
corresponding to nonzero eigen-
values using the singular value decomposition theorem.
3. Work out the matrix U of the frst r unit eigenvectors of ( )
T
B W
P S c S P - corre-
sponding to the largest eigenvalues.
4. Compute the discriminant matrix of MMSD using the formula V = PU.
Input: Training samples x
1
,x
2
,...,x
N
, class labels of these samples
1
( ),..., ( )
N
l l x x , parameter
value c, and the number of extracted features r
Output: The discriminant matrix of MMSD V
is prohibited.
all unit eigenvector of the total scatter matrix corresponding nonzero eigenvalues, P
using singular value decomposition theorem; second, we map the high-dimensional
input space into the range of the total scatter matrix using : ,
T d t T
P R R P x x ;
and third, we perform eigen decomposition on the matrix ( )
T
B W
P S c S P - . Algo-
rithm 6.1 is a detailed description of the novel facial feature extraction algorithm
based on MMSD.
6.1.3 t heoretical Analyses of MMSD Feature Extraction Method
To further investigate MMSD feature extraction method we explored its relationship
to other feature extraction approaches. Section 6.1.3.1 reveals the relation between
MMSD and Principal Component Analysis. Section 6.1.3.2 reveals the relation be-
tween MMSD and Karhunen-Love expansion. Section 6.1.3.3 reveals the relation
between MMSD and N-LDA. And Section 6.1.3.4 discusses the physical meaning
of the parameter of MMSD.
6.1.3.1 Connection to Principal Component Analysis
Although MMSD is a supervised feature extraction method, it is closely related
to a well-known unsupervised feature extraction approach, PCA. When the pa-
rameter c assumes the value of -1, the generalized scatter difference matrix is
( 1)
B W B W T
S S S S S - - = + = , i.e. the total scatter matrix. This implies that when c
= -1 MMSD discriminant vectors are in fact principal component directions.
In facial feature extraction, we can obtain eigenfaces of MMSD by reverting
MMSD discriminant vectors to images. Figure 6.1 displays the frst seven eigenfaces
of MMSD when the parameter c assumes the values of -1, 1, 10, 100, and 1000.
These eigenfaces are calculated on the training set, which consists of the frst fve
images of each individual from the ORL face image database. As shown earlier,
the eigenfaces of MMSD are actually Eigenface (Belhumeur et al., 1997) when the
parameter c assumes the value of -1 and as can be seen in Figure 6.1, the details in
the MMSD eigenfaces increases with the value of the parameter c.
6.1.3.2 Connection to K-L Expansion Based on the Between-Class
Scatter Matrix
When the parameter c assumes the value of 0, MMSD degenerates into a feature
extraction method which derives its discriminant vectors in the range of the be-
tween-class scatter matrix. In fact, MMSD is equivalent to the Karhunen-Love
(K-L) expansion whose generation matrix is the between-class scatter matrix S
B
.
Here discriminant vectors of MMSD are orthonormal eigenvectors of the matrix
is prohibited.
Figure 6.1. The frst seven eigenfaces of MMSD when the parameter c are -1, 1, 10,
100, and 1000, respectively
S
B
corresponding to the r largest eigenvalues. It implies that discriminant vectors
of MMSD are solely dependent on the objective (6.2) in this case.
6.1.3.3 Asymptotic Property of MMSD
The asymptotic property of MMSD is revealed by the following theorem.
Theorem 6.2 If S
W
is singular, discriminant vectors of MMSD are approaching
discriminant vectors of N-LDA when the value of the parameter c is approaching
infnity.
Proof: We only prove that the frst discriminant vector of MMSD is approaching
the frst discriminant vector of N-LDA when the value of the parameter c is ap-
proaching infnity. Other proofs are similar.
Let w
b
, w
c
be the unit eigenvectors of matrices S
B
,

(S
B
- c

S
W
) corresponding to
the largest eigenvalues
b
,
c
respectively. Since S
W
is a singular matrix, there ex-
is prohibited.
ists a nonzero unit vector w
0
such that S
W
w
0
= 0. Considering the fact that S
B
is a
semi-positive matrix, we have
0 0 0 0 0 0
1
max ( ) 0
T T T T
c b w b w b
S C S S C S S
=
= - - =
w
w w w w w w w w .
(6.7)
From the meaning of
c
and w
c
the following equation is always true for any
positive real number c:
( )
B W c c c
S c S - = w w . (6.8)
By combining equality (6.8), inequality (6.7) and the meaning of
b
we obtain:
1 1 1
( )
T T T
c W c c B c c c B c b
S S S
c c c
= - w w w w w w . (6.9)
Since the matrix S
W
is also semi-positive, the following inequality holds:
0
T
c W c
S w w . (6.10)
By combining inequalities (6.9) and (6.10) we can conclude that
lim 0
T
c W c
c
S
= w w . (6.11)
Thus we complete the proof of the theorem. From the theorem it is easy to un-
derstand that N-LDA is in fact an asymptotic formof MMSD.
6.1.3.4 The Physical Meaning of the Parameter
From previous discussions we fnd that, like C-LDA, MMSD derives its discrimi-
nant vectors both in the range of the between-class scatter matrix and in the null
space of the within-class scatter matrix. But MMSD is much more fexible than
C-LDA. The parameter c can be used to adjust the balance between the two sub-
spaces. When c = 0, the discriminant vectors of MMSD are solely from the range
of the between-class scatter matrix .With the increase of the value of c from zero
to infnite, the discriminant vectors of MMSD are more and more from the null
space of the within-class scatter matrix. When the parameter c is approaching
infnity, the discriminant vectors of MMSD are solely from the null space of the
within-class scatter matrix.
is prohibited.
MMSD was evaluated on the same subset of the FERET database as shown in
Section 3.4.4. Similarly, all images were resized to 8080 pixels and preprocessed
by histogram equalization. In all experiments we used the nearest neighbor (NN)
classifer with Euclidean distance.
6.1.4.1 Accuracy of MMSD vs. No. of Extracted Features and the
Parameter Value
Three images of each individual were randomly chosen for training, while the re-
maining four images were used for testing. Thus, the training sample set size was
600 and the testing sample set size was 800. In this way, we ran the system ten times
and obtained ten different training and testing sample sets. Figure 6.2 demonstrates
average recognition rates of MMSD over various numbers of extracted features and
different values of the parameter c. From Figure 6.2 we fnd two facts: First, the
effectiveness of MMSD is quite robust on the number of extracted features; second,
the effectiveness of MMSD is sensitive to the value of the parameter c.
6.1.4.2 Comparison of MMSD with Other Facial Feature Extraction
Methods
The experimental design is the same as in Section 6.1.4.1. To make N-LDA appli-
cable, PCA is frst used to compress a high-dimensional image space into the range
Figure 6.2. Average recognition rate of MMSD over the number of extracted features
and the value of the parameter c
19
49
79
109
139
169
199
1000
100
10
1
0.1
0.01
0.2
0.4
0.6
0.8
1
N
u
m
b
er
of
e
x
tra
cte
d
fe
a
tu
res
V
a
lu
e
o
f
th
e
p
a
r
a
m
e
te
r
c
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
is prohibited.
of the total scatter matrix. Figure 6.3 displays curves of average recognition rates
of MMSD (c = 10), NSM, D-LDA, Eigenface, and Fisherface. From Figure 6.3 we
fnd that MMSD is much more effective than the other four facial feature extraction
methods and Fisherface achieve its maximum value at 49.
6.1.4.3 Further Comparison of MMSD with Fisherface and Other
Methods
According to Yang et al. (2005) the accuracy of Fisherface is heavily dependent
on the number of principal components used in the PCA stage. In this subsection
we compare MMSD with N-LDA, D-LDA, and Fisherface in the cases where the
number of principal components used in the PCA stage varies from 50 to 400.
The experimental design is the same as in Section 6.14.1. The number of extracted
feature for each method is 49. The value of the parameter c of MMSD is 10. Fig-
ure 6.4 displays curves of average recognition rates of various feature extraction
methods over varying number of principal components. We can see that MMSD
outperforms the other three feature extraction methods when there are more than
100 principal components.
Taking into account the statement that the accuracy of D-LDA is heavily dependent
on the number of training samples per individual as mentioned in Section 5.2.3.4,,we
Figure 6.3. Average recognition rates of various methods vs. the number of extracted
features
is prohibited.
compare the effectiveness of MMSD with that of N-LDA, D-LDA, Eigenface,
Fisherface, and C-LDA when the number of training samples per individual var-
ies from 2 to 6. The experiment consisted of fve tests of seven runs each. In each
run of the ith test, (i + 1) images of each individual were used for training and the
remaining (6 - i) images for testing. Images of each individual numbered 1 to (i
+ 1), 2 to (i + 2), , 7 to i were used as training samples in the frst, second,...,
seventh run respectively. The number of extracted features for MMSD, N-LDA,
D-LDA, Fisherface, Eigenface, and C-LDA in the ith test are respectively 199, 199,
199, 199, 200i + 199, and 398. Figure 6.5 displays curves of average recognition
rates of various feature extraction methods over varying number of training samples
per individual. MMSD is again the most effective of the facial feature extraction
approaches that were tested.
6.1.4.5 Comparison of MMSD with MSD
The comparison of effectiveness of MMSD with effectiveness of MSD is conducted
on a small database, ORL (please refer to Section 3.2.3). In each of the ten runs,
we use fve images of each person for training and the remaining fve for testing.
The images of each person numbered 1 to 5, 2 to 6, , 10 to 4 are used as training
samples for the frst, second,, and the tenth run respectively. In the experiment,
the number of features extracted by MMSD is 39 and the classifer used in combi-
nation with MMSD is the nearest neighbor. Table 6.1 lists the means and standard
Figure 6.4. Average recognition rates of various methods vs. the number of prin-
cipal components
is prohibited.
deviations of recognition rates of MMSD and MSD using various values of the
parameter c. Experimental results indicate that MMSD is more effective than MSD
on the ORL face image database.
6.1.4.6 Ef.ciencies of MMSD and MSD on the ORL Database
The chief motivation for extending MSD to MMSD is to promote effciency. In this
subsection we compare the effciency of an MSD classifer with that of MMSD in
combination with a NN classifer on the ORL database. The experimental design
is the same as in Section 6.1.4.5. Figure 6.6 displays the time (seconds) taken by the
MSD classifer and MMSD in combination with a NN classifer for pattern clas-
Parameter c 10 100 1000
MMSD
Mean 0.9605 0.9625 0.9630
Std 0.0140 0.0153 0.0140
MSD
Mean 0.9505 0.9525 0.9515
Std 0.0192 0.0186 0.0197
Table 6.1. Means and standard deviations of recognition rates of MMSD and
MSD
Figure 6.5. Average recognition rates of various feature extraction methods vs. the
numbers of training samples per individual
is prohibited.
Figure 6.6. Comparison of the effciencies of MSD classifers with MMSD + NN
classifers on the ORL face image database
1 2 3 4 5 6 7 8 9 10
0
5
10
15
20
25
30
35
40
Number of tests
T
i
m
e

f
o
r

c
l
a
s
s
i
f
i
c
a
t
i
o
n
MSD
MMSD+NN
sifcation in each run on the ORL database. MMSD in combination with an NN
classifer is much faster than an MSD classifer.
6.1.4.7 Effciency Comparison of MMSD and Other Methods on the
FERET Database
In this subsection we compare the effciency of MMSD with that of N-LDA, D-
LDA, Fisherface, and C-LDA on the FERET database. The experimental design
is the same as in Section 6.1.4.4. Figure 6.7 displays the average time (seconds)
taken by various methods used for feature extraction in each test on the FERET
database. Although slower than D-LDA and Fisherface, MMSD is faster than N-
LDA and C-LDA.
6.2 f eAture extr Act Ion bAsed on coeff IcIents of
vAr IAnces
PCA and LDA are two popular feature extraction techniques in statistical pattern
recognition feld. It is generally believed that PCA extracts the most representative
features whereas LDA extracts the most discriminative features. Theoretically speak-
ing, LDA features should be more suitable for classifcation than PCA features but
PCA does outperform LDA for face recognition in certain circumstances (Martinez
& Kak, 2001).
is prohibited.
Extensive experimental results show that nullspace LDA (N-LDA) is one of
the most effective techniques for face recognition. N-LDA tries to fnd a set of
discriminant vectors which maximize the between-class scatter in the null space
of the within-class scatter matrix. The calculation of its discriminant vectors in-
volves performing singular value decomposition on a high-dimensional matrix. It
is generally memory- and time-consuming.
One of the key ideas of the N-LDA is to calculate discriminant vectors in the
null space of the within-class scatter matrix. Its main advantage is that projections
of training samples from each class are exactly the same when they are projected on
discriminant vectors. Thus, the difference between the projection of a test sample
and projections of training samples from a certain class is a good metric of its per-
tinence to this class. Borrowing the idea in N-LDA and the concept of coeffcient of
variance in statistical analysis we present a novel facial feature extraction method,
i.e. discriminant based on coeffcient of variance (DCV) in this section.
6.2.1 Dimensionality Reduction Based on Coeffcient of
variation
In this subsection, at frst we introduce the concept of the minimum coeffcient of
variance (MCV) direction, then present calculation methods for this direction, and
last propose a feature extraction method based on MCV directions for each class
of training samples.
Figure 6.7. Effciency comparison of MMSD with those of N-LDA, D-LDA, Fish-
erface, and C-LDA on the FERET database
2 3 4 5 6
0
50
100
150
200
250
300
350
400
A
v
e
r
a
g
e

t
i
m
e

f
o
r

f
e
a
t
u
r
e

e
x
t
r
a
c
t
i
o
n

(
s
e
c
.
)
MMSD
N-LDA
D-LDA
Fisherface
C-LDA
is prohibited.
6.2.1.1 Defnition of the MCV Direction
Before defning the concept of the MCV direction, let us briefy review the term of
coeffcient of variation of a dataset.
Let y
1
,...,y
n
be a set of scalars,
1
1
n
i
i
y y
n
=
=

and
2 2
1
1
( )
n
i
i
y y
n
=
= -
be its sample
mean and sample variance respectively. The coeffcient of variation (CV) of the
dataset y
1
,...,y
n
is defned as follows:
CV
y
=
. (6.12)
CV is a measure of scatter of a dataset. The smaller the CV score of a dataset
the less disperse the dataset.
Suppose that x
1
,...,x
n
is a group of d-dimensional samples from a certain class.
When they are projected on a one-dimensional subspace spanned by a direction
1 d
R

w ( w 0), a set of scalars
1
,...,
T T
n
w x w x is obtained and the corresponding
CV score, i.e. the CV score of the class of samples along the direction, CV(w) can
be calculated.
Defnition 6.1 A direction is called a MCV direction of a class of samples, if the
absolute value of the CV score of the set of samples from the class along this direc-
tion is the smallest.
The defnition of MCV direction means that when a d-dimensional dataset is
projected on its MCV direction the density of the projected dataset is the highest
and the square of the mean of the projected dataset is synchronously the largest. If
a direction is used to judge whether an unknown sample comes from a certain class
according to projections of sampleson it, the MCV direction of training samples
from the class is out of question the best choice.
Fig 6.8 shows the MCV direction and the PCA direction for a two-dimensional
dataset. From the fgure we fnd that: 1. While the PCA direction w
PCA
achieves the
maximum variance of the projected dataset, the MCV direction w
MCV
achieves the
minimum coeffcient of variance of the projected dataset; 2. The direction w
MCV

is slightly different from the direction w
0
which achieves the minimum variance
of the projected dataset and is perpendicular to the PCA direction w
PCA
; 3. We can
conclude that the sample x is an outlier of the dataset with enough confdence since
the projection of x on w
MCV
is apart from projections of samples from the dataset.
is prohibited.
6.2.1.2 Calculation of a MCV Direction
When dataset x
1
,...,x
n
is projected on a one-dimensional subspace spanned by a
direction w, the CV score of the dataset along this direction is
1
1
( )( )
( )
n
T T
i i
i
T
T
T
n
CV
S
=
(
- -
(

=
=
w x m x m w
w
w m
w w
w m
. (6.13)
Here
1
1
n
i
i
n
=
=

m x is the mean sample, and
1
1
( )( )
n
T
i i
i
S
n
=
= - -
x m x m the scatter
matrix. Thus, the MCV direction is a solution of the following optimization
model,
min
T
T
S w w
w m
, (6.14)
which is equivalent to the following one,
min
T
T T
S w w
w mm w
. (6.15)
From model (6.15) we fnd that if w
*
is a MCV direction, kw
*
is a MCV direction
as well. Here k is an arbitrary nonzero real number. Since they span the same one-
dimensional subspace they are viewed as one direction in this subsection. Thereafter
we assume that MCV directions are unit vectors for simplicity.
Theorem 6.3 and Theorem 6.4 give a solution of the optimization model (6.15)
when S is nonsingular or singular respectively.
Theorem6.3 If S is nonsingular, the eigenvector of the matrix
1 T
S
-
mm correspond-
ing to the largest eigenvalue is a MCV direction.
Proof. The Lagrangian Function of
( )
T
T T
S
f =
w w
w
w mm w
is
is prohibited.
( ) ( )
T T T
L S c = - - w w w w mm w
.
Let
0
L
=
w
, we have
1
1
T
S
-
= mm w w.
Suppose w
*
be an eigenvector of matrix
1 T
S
-
mm corresponding to the largest
eigenvalue . We have
* *
* *
1
T
T T
S
=
w w
w mm w
.
It means that the function f(w) achieves its minimum at w
*
. Thus, w
*
is a MCV
direction.
If S is singular, there might exist infnite different directions whose correspond-

ing CV values are zero. Thus, the MCV direction can not be properly calculated. In
this case, a MCV direction is defned as one unit vector which is in the null space
of S and maximizes the square of the mean of projected dataset. That is, a MCV
direction is a solution of the following optimization model:
( )
1
max
T T
N S
=
w
w
w mm w. (6.16)
Here ( ) { | }
d
N S R S = = x x 0 is the null space of S.
Theorem6.4 Suppose
1
,...,
k
be an orthonormal basis for N(S). The unit vector
/
T T
m m is a MCV direction. Here = [
1
,...,
k
]

, and
T
m is the 2-
norm of
T
m.
Proof For any vector w in N(S), there exists a group of scalars a
1
,...,a
k
such that
1
k
i i
i
a
=
= =
w a.
Here a = [a
1
,...,a
k
]
T
. Since
T
is an identity matrix, the optimization model (6.16)
can be transformed into:
1
max
T
T T T
=

a a
a mm a. (6.17)
By using Lagrange multiplier, it is easy to conclude that vector /
T T
= a m m
is one solution of above optimization problem. Thus, the unit vector /
T T
m m
is a MCV direction.
is prohibited.
When d, the dimensionality of a dataset, is very large the calculation of an
orthonormal basis for N(S) is not a simple task. It will consume a large amount of
memory and computing time to fnd a MCV direction. Handy calculating method
should be considered. Theorem 6.5 provides a simple approach to calculate a MCV
direction when S is singular.
Before presenting the theorem a lemma is introduced frst.
Lemma6.2 (Liu, Cheng, & Yang, 1992b) Let
n n
A R

be a real symmetric ma-
trix,
1 n
R

w a real column vector. It can be concluded that A = w 0 if and only if
T
A = w w 0.
Theorem6.5 A MCV direction for a dataset x
1
,...,x
n
can be calculated using the
formula
( ) / ( )
T T
MCV n n
X X
+ +
= w 1 1 .
Here X = x
1
,...,x
n
is the data matrix, A
+
is the Moore-Penrose inverse of A, 1
n
is
a column vector with all n elements being ones.
Proof: From the defnition of the scatter matrix S (
1
1
( )( )
n
T
i i
i
S
n
=
= - -
x m x m ), we
have
2
1
1
[ ( )] 0 [1,..., ], .
n
T T T T
i i
i
S i n
n
=
= - = =
w w 0 w x m w x w m
Since S is a real symmetric matrix, from Lemma 1 we know that
( )
T
N S S = w w w 0

[1,..., ], ( ) 1
T
i
T
i n =
w
x
w m
.
Let v denote /
T
w w m. The constrain of the optimization model (6.16) is converted
into
1, 1,...,
T
i
i n = = v x
and 1 = w .
Since the objective of the optimization model (6.16) is max
T
T T
T
=
w w
w mm w
v v
,
which can be rewritten as
2
min
T
v v
w
, the optimization model (6.16) can be trans-
formed into
is prohibited.
1, 1,...,
min
T
i
T
i n = = v x
v v, (6.18)
and
/ = w v v . (6.19)
The constraint of the optimization model (6.18) can be written in the follow-
ing matrix form
| | | |
1
1 1
T
n
= v x x which can be further converted into
T
n
X = v 1
. In a point of view of matrix theory the optimization model (6.18) tries
to fnd the minimum norm minimum square error solution for a system of linear
equations. Thus, the unique solution of the optimization model (6.18) can readily
be computed by
( )
T
n
X
+
= v 1
. (6.20)
6.2.1.3 Discriminant Based on Coeffcient of Variance

Assume x
1
,...,x
N
be a set of training samples from l classes, with N
i
samples from
the ith class. For class i (i = 1,,l) there is a MCV direction w
i
corresponding to
training samples from that class. Due to similarities between samples from different
classes, MCV vectors w
1
,...,w
l
are generally dependent. In order to eliminate linear
dependency among them, the Gram-Schmidt orthogonalization and normalization
procedure is performed here.
Let
1 1 1
/ , = u w w
(6.21)
1
1
( ) ,
i
T
i i j i j
j
-
=
= -
v w u w u
(6.22)
/ , 2 .
i i i
i l = u v v
(6.23)
We call u
1
,...,u
l
the discriminant vectors based on coeffcient of variance.
Once discriminant vectors u
1
,...,u
l
are calculated, the d-dimensional input space
can be mapped into a l-dimensional feature space by the discriminant matrix U =
u
1
,...,u
l
. We call this dimensionality reduction method discriminant based on coef-
fcient of variance (DCV). The feature extraction procedure of DCV is illustrated
in Figure 6.9. The physical meaning, advantages and disadvantages of DCV will
be discussed in some detail in Section 6.2.4.
is prohibited.
6.2.2 Experimental Evaluation
We applied DCV to face recognition in order to test its accuracy and compared
it with six famous facial feature extraction methods: Fisherface, the orthogonal
complimentary space method, Eigenface, N-LDA, uncorrelated linear discriminant
analysis (U-LDA), and D-LDA on two benchmark face image databases, i.e. the
FERET and AR (please refer to Section 3.4.4) face image database. The FERET
dataset is used to evaluate the performance of DCV under conditions where the
pose, sample size, and number of classes are varied. The AR dataset is employed
to test the performance of DCV under conditions where there is a variation in
facial expressions. In all experiments, the nearest neighbor decision rule is used
in classifcation.
Figure 6.9. Flowchart of DCV
Compute the
dataset for each
class
(X
1
,,X
l
)
Calculate the
corresponding
MCV direction
(w
1
,,w
l
)
Training
dataset
(X)
Perform Gram-Schmidt
orthogonalization and
normalization procedure
(u
1
,,u
l
)
Discriminant
matrix
(U = [u
1
,,u
l
])
Figure 6.8. Illustration of the MCV direction and the PCA direction for a two-di-
mensional dataset
is prohibited.
6.2.2.1 Experiment on the FERET Database
The same subset of the FERET database as shown in Section 3.4.4 is used in the
experiment. All images are also resized and pre-processed by histogram equaliza-
tion as shown in Section 3.4.4. In the experiment the number of classes, i.e. the
number of individuals, k is varied from 100 to 200 with an interval 10. For a given
number of classes there are seven tests for each feature extraction method. In each
test, three images of each person are used for training and the remaining four im-
ages for testing. Thus the total amount of training samples and testing samples are
3k and 4k respectively. Three training samples are taken in turn, i.e. face images
numbered 1 to 3, 2 to 4, , 7 to 2 for each individual are used as training samples
respectively.
Figure 6.10 illustrates the tendencies of average recognition rates of each facial
feature extraction method when the number of classes varies from 100 to 200.
Partial experimental results are shown in Table 6.2. From Table 6.2 and Figure 6.10
we fnd that: 1. As other facial feature extraction method, the average recognition
rate of DCV decreases with the increase of number of classes; 2. The proposed
method always outperforms Fisherface, Eigenface, U-LDA, and D-LDA; 3. DCV
is slightly inferior to N-LDA and OCS when the number of classes is small and is
superior to them when the number of classes is large enough. An interesting fact
is that OCS is as effective as N-LDA.
Figure 6.11 displays average times consumed by various feature extraction meth-
ods on this subset of FERET database. All results are derived from experiments
performed on a PC with a P4 3.0GHz CPU and 1GB main memory running Micro-
soft Windows XP Professional. Here PCA, N-LDA, and FLD stand for the feature
extraction methods Eigenface, nullspace LDA, and Fisherface respectively. From
Figure 6.11 we fnd that DCV is much more effcient than N-LDA and OCS.
6.2.2.2 Experiment on the AR Database
The 119 individuals each having 26 face images taken in two sessions of the AR
face database were involved in the experiment. For each individual, eight image
samples taken under different lighting conditions (Please refer to Figure 3.10 (a),
3.10 (e), 3.10 (f), 3.10(g), 3.10 (n), 3.10 (r), 3.10 (s), and 3.10 (t)) were employed in the
experiment. All images were also normalized and pre-processed using the same
approch as shown in Section 3.4.4. For each individual, a sample from the frst
section and a sample from the second section are used as training samples and the
remaining six samples are used for testing. Thus, the total numbers of training and
test samples are 240 and 720 respectively. Since there are 16 (= 44) choices in the
selection of training samples for each individual, the experiment contains 16 tests.
is prohibited.
Figure 6.10. Average recognition rates of various feature extraction methods vs. No.
of classes
100 110 120 130 140 150 160 170 180 190 200
0.3
0.4
0.5
0.6
0.7
0.8
Number of classes
A
v
e
r
a
g
e

r
e
c
o
g
n
i
t
i
o
n

r
a
t
e
DCV
Eigenface
N-LDA
D-LDA
U-LDA
OCS
Fisherface
Table 6.2. The average recognition rates and standard deviations with varying
number of classes on FERET database for various methods
Number of classes (k)
100 120 140 160 180 200
Fisherface
(k-1)
mean 0.4636 0.4128 0.3538 0.3174 0.2837 0.2518
std 0.1019 0.0960 0.0859 0.0896 0.0834 0.0833
OCS
(k-1)
mean 0.6486 0.6229 0.5901 0.5703 0.5478 0.5282
std 0.1051 0.1142 0.1220 0.1261 0.1296 0.1326
Eigenface
(3k-1)
mean 0.5779 0.5631 0.5515 0.5460 0.5317 0.5211
std 0.0664 0.0641 0.0725 0.0727 0.0716 0.0704
N-LDA
(k-1)
mean 0.6486 0.6229 0.5923 0.5703 0.5478 0.5282
std 0.1051 0.1142 0.1226 0.1261 0.1296 0.1326
U-LDA
(k-1)
mean 0.4957 0.4658 0.4291 0.4031 0.3865 0.3636
std 0.1017 0.0990 0.0956 0.0929 0.0891 0.1003
D-LDA
(k-1)
mean 0.5182 0.4911 0.4653 0.4379 0.4115 0.3914
std 0.1148 0.1084 0.1189 0.1118 0.1048 0.1068
DCV
(k)
mean 0.6200 0.6018 0.5842 0.5734 0.5595 0.5463
std 0.0718 0.0707 0.0724 0.0698 0.0667 0.0663
is prohibited.
Table 6.3 reports the average recognition rates and standard deviations of various
facial feature extraction methods on this subset of AR database.
From Table 6.3 we can conclude that DCV is superior to all feature extraction
methods except OCS and N-LDA. Figure 6.12 displays average times consumed
by various feature extraction methods on this subset of AR database. Here PCA
and FLD stand for Eigenface and Fisherface respectively. From Figure 6.12 we fnd
that DCV is again much more effcient than N-LDA and OCS.
6.2.3 discussion on dcv
6.2.3.1 Necessity for Orthnormalization of the Matrix of MCV
Vectors
In comparison to MCV vectors which are nature representatives for correspond-
ing individuals face images, DCV discriminant vectors emphasize deliberately
different parts of face images for different individuals. Figure 6.13 and 6.14 show
the two-dimensional displays of MCV vectors for some individuals in the FERET
database and their corresponding DCV discriminant vectors, respectively.
Figure 6.11. Average times for feature extraction of various methods on FERET
database
DCV PCA N-LDA D-LDA U-LDA OCS FLD
0
200
400
600
800
1000
1200
1400
1600
1800
2000
Feature extraction method
A
v
e
r
a
g
e

t
i
m
e

f
o
r

f
e
a
t
u
r
e

e
x
t
r
a
c
t
i
o
n

(
S
e
c
o
n
d
s
)
is prohibited.
Table 6.3. The average recognition rates and standard deviations of various methods
on a subset of AR database with variances in facial expression
Facial Feature Extraction Method
Fisherface
(119)
OCS
(119)
Eigenface
(239)
N-LDA
(119)
U-LDA
(119)
D-LDA
(119)
DCV
(120)
mean 0.8761 0.9477 0.9189 0.9477 0.9082 0.2244 0.9270
std 0.0272 0.0174 0.0204 0.0174 0.0206 0.0174 0.0192
Figure 6.12. Average times for feature extraction of various methods on AR data-
base
DCV PCA N-LDA D-LDA U-LDA OCS FLD
0
100
200
300
400
500
600
Feature extraction method
A
v
e
r
a
g
e

t
i
m
e

f
o
r

f
e
a
t
u
r
e

e
x
t
r
a
c
t
i
o
n

(
S
e
c
o
n
d
s
)
6.2.3.2 General Properties of the DCV Transform
Similar to LDA transforms, the DCV transform maps high-dimensional input spaces
into low-dimensional feature spaces such that the samples from the same classes
are as close as possible and the samples from the different classes are as apart
as possible. Unlike most LDA algorithms which synchronously achieve the two
goalsminimizing the within-class distances and maximizing the between-class
distances, DCV achieves these two goals in two phases: First, it tries to minimize
the within-class distances by projecting samples along MCV directions; second, it
tries to maximize the between-class distances by eliminating the linear dependences
among the MCV vectors. The discrimination process of DCV quite resembles that
is prohibited.
of human beings. When we discriminate different objects we frst try to capture the
chief characteristics of each object and then attempt to distinguish the differences
between these characteristics.
Since the DCV calculates its discriminant vectors in two phases, it is much scal-
able for large-scale face recognition tasks in the following sense: When samples
from new classes successively pool in the training dataset there is no need to recal-
culate all discriminant vectors but the MCV vectors and the corresponding DCV
discriminant vectors of the newly added classes.
6.3 summAr y
In this chapter we introduce two novel facial feature extraction methods. The frst
is multiple maximum scatter difference (MMSD) which is an extension of a bi-
nary linear discriminant criterion, i.e. maximum scatter difference. The second is
discriminant based on coeffcients of variances (DCV) which can be viewed as a
generalization of N-LDA. Experimental studies demonstrate that these two methods
are both promising. Especially, MMSD outperforms almost all state-of-the-art facial
feature extraction methods in terms of recognition accuracy.
Figure 6.14. Two-dimensional displays of the DCV discriminant vectors correspond-
ing to above individuals
10
20
30
40
50
60
70
80
10
20
30
40
50
60
70
80
Figure 6.13. Two-dimensional displays of MCV vectors for some individuals in
FERET face image database
is prohibited.
r eferences
Belhumeur, P. N., Hespanha, J. P., & Krieqman, D. J. (1997). Eigenfaces vs. Fisher-
33(10), 1713-1726.
Fukunaga, K. (1990). Introduction to statistical pattern recognition (2nd ed.).
Boston: Academic Press.
Guo, Y. F., Li, S. J., Yang, J-Y., Shu, T. T., & Wu, L. D. (2003). A generalized Foley-
Sammon transform based on generalized Fisher discriminant criterion and its ap-
plication to face recognition. Pattern Recognition Letters, 24(1-3), 147-158.
Hong, Z. Q., & Yang, J-Y. (1991). Optimal discriminant plane for a small number of
samples and design method of classifer on the plane. Pattern Recognition, 24(4),
317-324.
Hsu, C., & Lin, C. (2002). A comparison of methods for multiclass support vector
machines. IEEE Trans. Neural Networks, 13(2), 415-425.
Liu, K., Cheng, Y. Q., & Yang, J-Y. (1992a). A generalized optimal set of discrimi-
nant vectors. Pattern Recognition, 25(7), 731-739.
Liu, K., Cheng, Y. Q., & Yang, J-Y. (1992b). An effcient algorithm for Foley-Sam-
mon optimization set of discriminant vectors by algebraic method. International
Martinez, A. M., & Kak, A. C. (2001). PCA versus lda. IEEE Trans. Pattern Anal.
Song, F. X., Liu, S. H., & Yang, J-Y. (2006). Face representation based on the
multiple-class maximum scatter difference. Acta Automatica Sinica (in Chinese),
32(3), 378-385.
165-174.
Song, F. X., Zhang, D., Mei, D. Y., & Guo, Z. W. (2007). A multiple maximum
scatter difference discriminant criterion for facial feature extraction. IEEE Trans.
System, Man and Cybernetics, Part B, 37(6), 1599-1606.
is prohibited.
Tao, D. C., Li, X. L., Hu, W. M., Maybank, S. J., & Wu, X. D. (2005). Supervised
tensor learning. Proc. IEEE Intl Conf. Data Mining (pp. 450-457).
Tao, D. C., Li, X. L., Maybank, S. J., & Wu, X. D. (2006). Human carrying status
in visual surveillance. Proc. IEEE Conf. Computer Vision and Pattern Recognition
(pp. 1670-1677).
Foley- Sammon transform. Optical Engineering, 25(7), 834-840.
Turk, M., & Pentland, A. (1991). Face Recognition Using Eigenfaces. Proc. IEEE
Conf. Computer Vision and Pattern Recognition (pp. 586-591).
Yang, J., Frangi, A. F., Yang, J-Y., Zhang, D., & Jin, Z. (2005). KPCA plus lda: a
complete kernel Fisher discriminant framework for feature extraction and recogni-
tion. IEEE Trans. Pattern Anal. Machine Intell., 27(2), 230-244.
Yu, H., & Yang, J. (2001). A direct LDA algorithm for high-dimensional data with
application to face recognition. Pattern Recognition, 34 (10), 2067-2070.
Zhao, W., Chellappa, R., & Phillip, P. J. (1999). Subspace linear discriminant analysis
for face recognition (Technical Report CAR-TR-914). Baltimore, MD: Center for
Automatic Research, University of Maryland.
Section II
Tensor Technology
Tensor Space 135
is prohibited.
Chapter VII
Tensor Space
Abstr Act
In this chapter, we frst give the background materials for developing tensor
discrimination technologies in Section 7.1. Section 7.2 introduces some basic
notations in tensor space. Section 7.3 discusses several tensor decomposition
methods. Section 7.4 introduces the tensor rank.
7.1 bAckground
Matrix decompositions, such as the singular value decomposition (SVD), are ubiqui-
tous in numerical analysis. One usual way to think of the SVD is that it decomposes
a matrix into a sum of rank-1 matrices. In other words, an I
1
I
2
matrix A can be
expressed as a minimal sum of rank-1 matrices:
A = u
1
v
1
+ u
2
v
2
+
...
+ u
r
v
r
, (7.1)
where u
i

R
I
1
and v
i

R
I
2
for all i = 1,
...
,r. The operator denotes the outer product.
Thus the ijth entry of the rank-1 matrix a b is the product of the ith entry of a and
the jth entry of b, denoted by (a b)
ij
= a
i
b
j
. Such decompositions provide possibilities
to develop fundamental concepts such as the matrix rank and the approximation
theory and gain a range of applications including WWW searching and mining,
image processing, signal processing, medical imaging, and principal component
analysis. The decompositions are well-understood mathematically, numerically,
and computationally.
is prohibited.
A tensor is a higher order generalization of a vector or a matrix. In fact, a vector
is a frst-order tensor and a matrix is a tensor of order two. Furthermore speaking,
tensors are multilinear mapping over a set of vector spaces. If we have data in three
or more dimensions, then we mean to deal with a higher-order tensor. In tensor
analysis, higher-order tensor (also known as multidimensional, multiway, or n-way
array) decompositions (Martin, 2004; Comon, 2002) are used in many felds and
also have received considerable theoretical interest.
Different from some classical matrix decompositions, extending matrix decom-
positions such as the SVD to higher-order tensors has proven to be quite diffcult.
Familiar matrix concepts such as rank become ambiguous and more complicated.
One goal of the tensor decomposition is the same as for a matrix decomposition:
to rewrite the tensor as a sum of rank-1 tensors. Consider, for example, an I
1
I
2

I
3
tensor A. We would like to express A as the sum of rank-1 third-order tensors,
that is,
A = u
1
v
1
w
1
+ u
2
v
2
w
2
+
...
+ u
r
v
r
w
r
, (7.2)
where, u
i
R
I
1
, v
i
R
I
2
, and w
i
R
I
3
for all i = 1,
...
, r. Note that if a, b, c are vec-
tors, then (a b c)
ijk
= a
i
b
j
c
k
.
7.2 bAsIc not At Ions
In this section, we introduce some elementary notations and defnitions needed in
the later chapter.
If A is an I
1
...
I
N
tensor, then the order of A is N. The nth mode, way, or dimension
of A is of size I
n
(n = 1,
...
, N).
7.2.1 t ensors as vectors
First, let us defne the operators. Let B R
I
1
I
2
. Then vec(B) is defned as
1 2
2
(:,1)
( )
(:, )
I I
B
vec B
B I
(
(
=
(
(

. , (7.3)
where B(:,i) denotes the ith column of matrix B.
In other words, the vec operator is to transform a matrix into a vector by stack-
ing columns of matrix B.
Tensor Space 137
is prohibited.
Let b R
I
1
I
2
. Then reshape(b, I
1
I
2
is defned as
reshape(b, I
1
I
2
=
1 1 1 2 1 1 2
(1: ( 1: 2 ) (( 1) 1: b I b I I b I I I I
(
+ - +

R
I
1
I
2
.
(7.4)
In other words, reshape(b, I
1
I
2
produces an I
1
I
2
matrix from b. In fact, the
operators vec and reshape are related. It is straightforward to obtain
b=vec(reshape(b, I
1
I
2
), (7.5)
vec(B)=reshape(B, I
1
I
2
, 1. (7.6)
Similarly, for,A R
I
1
I
2
I
3
then vec(A) is defned as
1 2 3
3
( (:,:,1))
( )
(:,:, )
I I I
vec A
vec A
vec I

(
(
=
(
(

. . (7.7)
For example: for A R
I
1
I
2
I
3
, I
1
= I
2
= I
3
= 2
111
211
121
221
112
212
122
222
( )
a
a
a
a
vec A
a
a
a
a
(
(
(
(
(
(
=
(
(
(
(
(
(

.
7.2.2 t ensors as matrices
In real applications, it is often necessary to represent tensors as matrices. Typically,
all the columns along a certain mode are rearranged to form a matrix. Transforming
a tensor into a matrix in this way is referred as fattening, matricizing, or matrix
unfolding. In general, there are multiple ways to order the columns. More than
often not, the mode-n matricizing or matrix unfolding of an Nth-order A (I
1
...
I
N
)
is the set of vectors in R
I
n
obtained by keeping the index I
n
fxed and varying the
is prohibited.
other indices. Therefore, the mode-n matricizing of an Nth-order tensor is a matrix
( )
n n
I I
n
A

, where

n i
i n
I I
.
In other words, the matrix unfolding
1 1 1
( )
( )
n n N n
I I I I I
n
A
+ -

contains the
elements 1 N
i i
a

at the position with row number i
n
and column number equal
t o
1 2 1 1
( 1)
n n N n
i I I I I
+ + -
-
2 3 1 1
( 1)
n n N n
i I I I I
+ + -
+ - + +
1 1
( 1)
N n
i I I
-
-
1 2 1
( 1)
n
i I I
-
+ -
2 3 1 1
( 1)
n n
i I I i
- -
+ - + + .
For a third-order tensor, there are three possible fattening matrices. Figure 7.1
shows three possible cuts for a 2 2 2 tensor.
As a result, there are three possible fattening matrices for a 2 2 2 tensor A,
denotes A as follows:
Figure 7.1. An illustration of a 2 2 2 tensor that is cut in three ways
Tensor Space 139
is prohibited.
7.2.3 Norms and inner Products, Outer Product
If A, B R
I
1
I
2
I
3...
I
N
, then the inner product is
3 1 2
1 1
1 2 3
1 1 1 1
,
N
N N
N
I I I I
i i i i
i i i i
A B a b
= = = =
=

= vec(A)
T
vec(B). (7.8)
If x R
I
1
,

y R

I
2
,
then the outer product,
T
yx , is a rank-1 matrix. Note that
( )
T
vec yx x y . (7.9)
More generally, if x, y, z are vectors, x y z is a rank-1 tensor.
If A = UV
T
, U = [u
1
,
...
u
n
], and V = [v
1
,
...
v
n
], then
1 1
n n
T
ij i j
i j
A u v
= =
=
. (7.10)
Figure 7.2. Three fattening matrices of a third-order tensor

A=
is prohibited.
Further, it is easy to verify
1 1
( )
n n
ij j i
i j
vec A v u
= =
=
= V U)vec(). (7.11)
When A is a three-order tensor, we have
1 1 1
( )
n n n
ijk k j i
i j k
vec A w v u
= = =
=
.
=
( ) ( ) W V U vec
. (7.12)
where U = [u
1
,
...
,u
n
], V = [v
1
,
...
,v
n
], and W = [w
1
,
...
,w
n
].
From Eq. (7.12), we have
A
(1)
= U
(1)
(V W)
T
,
A
(2)
= V
(2)
(W U)
T
,
A
(3)
= W
(3)
(U V)
T
.
7.2.4 N-Mode Multiplication
To multiply a tensor by a matrix, one needs to specify which mode of the tensor
is multiplied by the columns of the matrix. Let us frst have a look at the matrix
product G = U F V
T
, involving three matrices F R
I
1
I
2
, U R
J
1
I
1
and V
R
J
2
I
2
. Note that the relationship between U and Fand the relationship between V
and F are completely similar: in the same way that U makes linear combinations
of the rows of F, V makes linear combinations of the column of the column of F. In
other words, the columns of F are multiplied by U and the rows of F are multiplied
by V. Moreover, one can see that the columns of U are associated with the column
space of G and the columns of V are associated with the row space of G. This typical
relationship will be denoted by means of the
n
symbol G = F
1
U
2
V.
In general, the following defnition (Lathauwer, Moor, & Vandewalle, 1994,
1996a, 1996b)for the n-mode product of a tensor is given.
Defnition 7.1. The n-mode product of a tensor A R
I
1
I
2
I
3...
I
N
by a matrix U
R
J
n
I
n
, denoted by A
n
U is an (I
1
I
2
...
I
n-1
J
n
I
n+1
...
I
N
)-tensor of which the
entries are given by
Tensor Space 141
is prohibited.
1 1 1
( )
n n n N
n i i j i i
A U
- +

1 1 1 n n n N n n
n
i i i i i j i
i
a u
- +
=

. (7.13)
Based on the defnition of the n-mode product, one can obtain
Property7.2. Given the tensor A R
I
1
I
2
I
3...
I
N
and the matrices F R
J
n
I
n
and
G R
J
m
I
m
, one has
(A
n
F)
m
G = (A
m
G)
n
F = A
n
F)
m
G. (7.14)
Property7.3. Given the tensor A R
I
1
I
2
I
3
...

I
N
and the matrices F R
J
n
I
n
and
G R
K
n
I
n
, one has
(A
n
F)
n
G = (A
n
(G F). (7.15)
Often the following simplifed notations are used in the case of the n-mode
product of a tensor by some researchers.
1 1 2 2
1
N
def
N N k k
k
A U U U A U
=
=
, (7.16)
and
1 1 2 2 1 1 1 1
1,
N
def def
i i i i N N k k i i
k k i
A U U U U U A U A U
- - + +
=
= =
,

(7.17)
where U
i
(i = 1,
...
,N) are matrices.
Figure 7.3 shows the equation A = B
1
U
1
2
U
2
3
U
3
for the third-order tensors
A R
J
1
J
2
J
3
, and B R
I
1
I
2
I
3
. Different from the classical way to visualize the
second-order matrix product, U
2
has not been transposed for reasons of symmetry.
Multiplication with U
1
needs linear combinations of the horizontal matrices (index
i
1
fxed) in B. Multiplication of B with U
1
means that every column of B (index i
2

and i
3
) has to be multiplied from the left with U
1
. In a similar way, multiplication
with U
2
needs linear combinations of the horizontal matrices (index i
2
fxed) in B
and multiplication with U
3
needs linear combinations of the horizontal matrices
(index i
3
fxed) in B. Figure 7.3 further provides the insight for understanding the
tensor technology. In fact, the n-mode product of a tensor and a matrix is a special
case of the inner product in multilinear algebra and tensor analysis. In the previous
literature, it is often used in the form of an Einstein summation.
is prohibited.
7.3 t ensor decom Pos It Ion
There are two types of decompositions used most in real applications. The frst
decomposition is the CANDECOMP-PARAFAC (CP) model (Carroll & Chang,
1970; Harshman, 1970) and the second decomposition is refereed to as the TUCKER
model (Tucker, 1966). These two models are widely used in three-order tensors and
in fact they can be extended to arbitrary ordered tensors.
7.3.1 cAndecom P-PAr Af Ac decomposition
For the sake of simplicity, let us consider three-order tensors. Given an I
1
I
2
I
3

tensor, A, the CP model has a decomposition of the form
1
R
i i i
i
A u v w
=
=
, (7.18)
where u
i
R
I
1
, v
i
R
I
2
and w
i
R
I
3
for i = 1,
...
,R.
Without loss of generality, the vectors are assumed to be real. However, the model
is also valid for the complex-valued vectors. Note that there are no constraints (such
as orthogonality) on the vectors u
i
R
I
1
, v
i
R
I
2
and w
i
R
I
3
. However, one can

1
J

A
2
J
3
J
=
1
I

B
2
I
3
I
1
J
1
U
1
I
2
J
2
U
2
I
3
J
3
U
3
I
Figure 7.3. Visualization of the multiplication of a third tensor B R
J
n
I
n
B
R
I
1
I
2
I
3
with matrices U
1
R
J
1
I
1
, U
2
R
J
2
I
2
and U
3
R
J
3
I
3
.
Tensor Space 143
is prohibited.
impose the constraints such as orthogonality, nonnegativity, or unimodality when
needed. Note that a CP decomposition always exists (take R to be the product of
the sizes of each mode and take outer products of scaled standard basis vectors).
Ideally, R is chosen to be the minimal number of terms needed to sum to A. When
R is minimal, then R is known as the tensor rank and is discussed in Section 7.4.
7.3.2 tucker decomposition
In this subsection, we frst consider three-order tensors. Given an I
1
I
2
I
3
tensor,
A, the TUCKER model has a decomposition of the form
3 1 2
1 1 1
( )
R R R
ijk i i i
i j k
A u v w
= = =
=
, (7.19)
where R
1
I
1
, R
2
I
2
, R
3
I
3
,u
i
R
R
1
,v
i
R
R
2
and w
i
R
R
3
for i, j, k.
The tensor, ( )
ijk
S = , is called the core tensor. Note that the core tensor does
not always need to have the same dimensions as A. From Eq. (7.19),one can see
that the CP decomposition is a special case of the TUCKER decomposition. Note
that there are no constraints on the vectors u
i
R
I
1
, v
i
R
I
2
and w
i
R
I
3
in the
TUCKER decomposition. However, one may impose constraints when needed. If
the u
i
R
I
1
, v
i
R
I
2
and w
i
R
I
3
are columns from orthogonal matrices U,V,W,
then the TUCKER model is referred to as the Higher-Order Singular Value De-
composition, or HOSVD.
Futher research showed Higher-Order Singular Value Decomposition by the
following theorem (Lathauwer, 1997; Lathauwer, Comon, Moor, & Vandewalle,
1995; Lathauwer, Moor, & Vandewalle, 1997, 2000).
Theorem 7.4. (nth order SVD). Every A R
I
1
I
2
I
3...
I
N
can be written as the
product
A = S
1
U
1
2
U
2
...
N
U
N
(7.20)
in which
1.
1
( , )
n
n n
n I
U u u = is a unitary I
n
I
n
matrix,
2. S is an I
1

...
I
N
tensor of which the subtensor
n
i
S
=
,obtained by fxing the
nth index to , have the properties of
(i) all-orthogonality: two subtensors
n
i
S
=
and
n
i
S
=
are orthogonal for all
possible values of n, and subject to : <
n
i
S
=
,
n
i
S
=
>=0 when
.
is prohibited.
(ii) ordering: ||
1
n
i
S
=
||
2
|| ||
n
i
S
=
|| || 0
n n
i I
S
=
for all possible values of n.
The Frobenius-norm || || 0
n
i i
S
=
, symbolized by
n
i
are n-mode singular values
of A and the vector
n
i
u is an ith n-mode singular vector.
For a tensor A R
I
1
I
2
I
3
,Figure 7.4 shows HOSVD for a third-order tensor.
Theorem 7.4 says that it is always possible to fnd orthogonal transformations of the
column, row, and 3-mode space such that S = A
1
(U
1
)
T
2
(U
2
)
T
3
(U
3
)
T
is all-orthogo-
nal and ordered. Note that the new basis vectors are the columns of U
1
, U
2
and U
3
.
All-orthogonality means that the different horizontal matrices of S (the frst index
i
1
kept fxed, while the two other indices, i
2
and i
3
are free) are mutually orthogonal
with respect to the scalar product of matrices (i.e., the sum of the products of the
corresponding entries vanishes). At the same time, the different frontal matrices
(i
2
fxed) and the different vertical matrices (i
3
fxed) should be mutually orthogonal
as well. The ordering constraint imposes that the Frobenius-norm of the horizontal
(frontal, resp., vertical) matrices does not increase as the index i
1
(i
2
, resp., i
3
) is
increased. While the orthogonality of U
1
, U
2
and U
3
, and the all-orthogonality of S
are the basic assumptions of the model, the ordering condition should be regarded
as a convention, meant to fx a particular ordering of the columns of U
1
, U
2
and U
3

(or the horizontal, frontal, and vertical matrices of S, stated otherwise).
There are some similar characteristics between the matrix decomposition and the
tensor decomposition. First, the role of the singular values in tensor decomposition
is taken over by the Frobenius-norms of the (N-1)th-order subtensors of the core
tensor S . Second, the left and right singular vectors of a matrix are generalized as
the n-mode singular vectors in tensor decomposition. Note at this point that in the
matrix case, the singular values also correspond to the Frobenius-norms of the rows
and the columns of the core matrix S. For Nth order tensors, N (possibly different)
sets of n-mode singular values are defned. In this respect, an Nth-order tensor can

1
I
A
2
I
3
I
=
1
U
1
I
1
I
1
I

B
2
I
3
I
3
I
3
U
3
I
2
I
2
U
2
I
Figure 7.4. Visualization of HOSVD for a third-order tensor
Tensor Space 145
is prohibited.
also have N different n-rank values. The essential difference is that S is in general
a full tensor, instead of being pseudodiagonal (this would mean that nonzero ele-
ments could only occur when the indices i
1
= i
2
=
...
i
N
). Instead, S obeys the condi-
tion of all-orthogonality. Here notice that in the matrix case S is all-orthogonal as
well due to the diagonal structure and the scalar product of two different rows or
columns also vanishes. We also observe that, by defnition, the n-mode singular
values are positive and real, like in the matrix case. On the other hand the entries
of S are not necessarily positive in general. They can even be complex, when A is
a complex-valued tensor.
In addition, we say that a tensor is diagonalizeable if the HOSVD yields a diagonal
core tensor. Note that if the core is diagonal, we can write the HOSVD as a PARA-
FAC with orthogonality constraints. In general a tensor cannot be diagonalized. Up
till now, it is not clear under what conditions permit a diagonalizeable core.
7.4 t ensor rA nk
Tensor rank (Kruskal, 1989; Hastad, 1990) is also an important problem in tensor
analysis. There are major differences between matrices and higher-order tensors
when rank properties are concerned. The differences may directly affect the way
that an SVD generalization could look. As a matter of fact, there is not a unique
way to generalize the rank concept in general. First, it is easy to generalize the
notion of column and row rank. If we refer in general to the column and row vec-
tors of an Nth-order tensor A R
I
1
I
2
I
3
...

I
N
as its n-mode vectors, defned as the
I
n
-dimensional vectors obtained from A by varying the index i
n
and keeping the
other indices fxed, then the following defnition is given.
Defnition 7.5. The n-rank of A, denoted by R
n
= rank
n
(A), is the dimension of the
vector space spanned by the n-mode vectors.
The n-rank of a given tensor can be analyzed by means of matrix techniques.
Property7.6. The n-mode vectors of A are the column vectors of the matrix un-
folding A
(n)
and
rank
n
(A) = rank(A
(n)
). (7.21)
A major difference with the matrix case, however, is the fact that the different
n-ranks of a higher-order tensor are not necessarily the same, as can easily be veri-
fed by checking some examples in the following.
is prohibited.
The rank of a higher-order tensor is usually defned in analogy with the fact that
a rank-R matrix can be decomposed as a sum of R rank-1 terms.
Defnition 7.7. An Nth-order tensor A has rank 1 when it equals the outer product
of N vectors, i.e.
1 1
1
N N
N
i i i i
a u u =
, (7.22)
for all values of the indices.
The rank of an arbitrary Nth-order tensor A, denoted by R =rank(A), is the
minimal number of rank-1 tensors that yield A in a linear combination.
From the defnition of a rank-1 tensor, a remark on the notation has to be made.
For matrices, the vector-vector outer product of a and b is denoted as
T
ab . To avoid
an ad hoc notation based on generalized transposes, one will further denote the
outer product of u
1
,
...
u
N
by u
1

...
u
N
.
Another difference between matrices and higher-order tensors is the fact that the
rank is not necessarily equal to an n-rank, even when all the n-ranks are the same.
From the defnitions it is clear that we always have R
n
R . In general, there is no
known method to compute tensor rank. It is shown in Hastad (1990) that tensor
rank is NP-complete.
In the following, some examples are shown with the context of comparing matrix
rank to tensor rank and complexity of understanding tensor rank.
Example7.1Let the feld of scalars be R, and suppose that
Then rank(A)=2:
1 1 1 1 1 1
1 1
A
1 1 1 1 1 1 2 2
( ( ( ( ( (
= +
( ( ( ( ( (
- - -

Tensor Space 147
is prohibited.
Example7.2 Let the feld of scalars be R, and suppose that
Then rank(A)=3:
.
1 1 1 1 1 1 1 1 1
1 1
A 2
1 1 1 1 1 1 0 0 0 2 2
( ( ( ( ( ( ( ( (
= + -
( ( ( ( ( ( ( ( (
- - -

Example7.3 Let the feld of scalars be C, and suppose that
Then rank(A)=2:
1 1
1 1
A
1 1 1 1 2 2
i i i i
i i
- -
( ( ( ( ( (
= +
( ( ( ( ( (
-

.
Examples 7.2 and 7.3 illustrate that the rank depends on the feld of scalars. In
general, the maximum rank of a 222- tensor over R is 3 but the maximum rank
over C is 2.
Example7.4 Consider the (222)-tensor A defned by
111 221 112
211 121 212 122 222
1
0
a a a
a a a a a
= = =
= = = = =
,
It follows that R
1
= R
2
= 2 but R
3
= 1.
is prohibited.
Example7.5.Consider the (222)-tensor A defned by
211 121 112
111 122 212 221 222
1
0
a a a
a a a a a
= = =
= = = = =
The 1-rank, 2-rank, and 3-rank are equal to 2. The rank, however, equals 3,
since
2 1 1 1 2 1 1 1 2
A X Y Z X Y Z X Y Z = + + .
111 221 112
211 121 212 122 222
1
0
a a a
a a a a a
= = =
= = = = =
in which
1 1 1
1
0
X Y Z
| |
= = =
|
\ .
,
2 2 2
0
1
X Y Z
| |
= = =
|
\ .
.
is a decomposition in a minimal linear combination of rank-1 tensors.
Example7.6. Let the feld of scalars be R, and suppose that
Then rank(A)=2:
1 1 1 1 1 1
A
1 1 1 0 0 0
( ( ( ( ( (
= +
( ( ( ( ( (

.
Example 7.6 shows that the minimal tensor decomposition is not always or-
thogonal. In other words, it is impossible to write A as the sum of two orthogonal
tensors (i.e., u
1
u
2
v
1
v
2
w
1
w
2
).
In addition, if the feld of scalars is R, then the following results are known:
(a) The maximum rank of an 222 tensor is 3,
(b) The maximum rank of an 333 tensor is 5,
(c) The maximum rank of an 888 tensor is 11.
r eferences
Carroll, J. D., & Chang, J. (1970). Analysis of individual differences in multidi-
mensional scaling via an n-way generalization of Eckart-Young decomposition.
Psychometrika, 35(3), 283-319.
Tensor Space 149
is prohibited.
Comon, P. (2002). Tensor decompositions. In J. G. McWhirter & I. K. Proudler
(Eds.), Mathematics in signal processing V. Oxford: Clarendon Press.
Harshman, R. A. (1970). Foundations of the PARAFAC procedure: model and
conditions for an explanatory multi-mode factor analysis. UCLA Working Papers
in phonetics, 16(1), 1-84.
Hastad, J. (1990). Tensor rank is NP-complete. J. Algorithms, 11(4), 644-654.
Kruskal, J. B. (1989). Rank, decomposition, and uniqueness for 3-way and n-way
arrays. In R. Coppi & S. Bolasco (Eds.), Multiway data analysis. Amsterdam:
Elsevier.
Lathauwer, L. (1997). Signal processing based on multilinear algebra. Ph.D thesis,
K.U. Leuven, Belgium.
Lathauwer, L., Comon, P., Moor, B., & Vandewalle, J. (1995). Higher-order power
method-application in independent component analysis. Proceedings of the Inter-
national Symposium on Nonlinear Theory and Its Applications (pp. 91-96).
Lathauwer, L., Moor, B., & Vandewalle, J. (1994). Blind source separation by higher
order singular value decomposition. Proc. of the 7th European Signal Processing
Conference (pp. 175-178).
Lathauwer, L., Moor, B., & Vandewalle, J. (1996a). Blind source separation by
simultaneous third-order tensor diagonalization. Proc. of the 8th European Signal
Processing Conference (pp. 2089-2092).
Lathauwer, L., Moor, B., & Vandewalle, J. (1996b). Independent component analysis
based on higher-order statistics only. Proceedings of the IEEE Signal Processing
Workshop on Statistical Signal and Array Processing (pp. 356-359).
Lathauwer, L., Moor, B., & Vandewalle, J. (1997). Dimensionality reduction in
higher order only ICA. Proceedings of the IEEE Signal Processing Workshop on
HOS (pp. 316-320).
Lathauwer, L., Moor, B., & Vandewalle, J. (2000). A multilinear singular value
decomposition. J. Matrix Anal. Appl., 21(4), 1253-1278.
Martin, C. D. (2004). Tensor decompositions workshop discussion notes. American
Institute of Mathematics (AIM).
Tucker, L. R. (1966). Mathematical notes of three-mode factor analysis. Psy-
chometrika, 31(3), 279-311.
is prohibited.
Chapter VIII
Tensor Principal Component
Analysis
Abstr Act
Tensor principal component analysis (PCA) is an effective method for data re-
construction and recognition. In this chapter, some variants of classical PCA are
introduced and the properties of tensor PCA are analyzed. Section 8.1 gives the
background and development of tensor PCA. Section 8.2 introduces tensor PCA.
Section 8.3 discusses some potential applications of tensor PCA in biometrics.
Finally, we summarize this chapter in Section 8.4.
8.1 Introduct Ion
Principal component analysis (Turk & Pentland, 1991; Penev & Sirovich, 2000),
also known as Karhunen-Love (K-L) transform, is a classical statistical technique
that has been widely used in various felds, such as face recognition, character
recognition, and knowledge representation. The aim of PCA is to reduce the dimen-
sionality of the data so that the extracted features are representative as possible. In
general, the key idea of PCA is to project data to an orthogonal subspace, which
can transform correlated variables into a smaller number of uncorrelated variables.
The frst principal component can capture variance of the data along some direc-
tion as possible, and consequent components capture as much of the remaining
variability as possible. Up to now, there are a number of theoretical analyses and
Tensor Principal Component Analysis 151
is prohibited.
discussion for PCA in the literature and PCA is one of the most popular methods
for data representation.
In recent years, some researchers (Yang, Zhang, Frangi, & Yang, 2004; Ye,
2004) noted that classical PCA often runs up against computational limits due to the
high time and space complexity for dealing with large image matrices, especially
for images and videos. In applying PCA, data must be converted to a vector form.
This results in the diffculty in eigen-decomposition in a high dimensional vector
space. To overcome this limitation, a novel idea is developed. This novel idea lies
in dealing with image matrices or video data directly rather than converting them
into vectors prior to dimensionality reduction. Based on this, Yang, Zhang, Frangi
and Yang (2004) proposed a two-dimensional PCA for image representation, whose
idea is that 2D image matrices are used to directly construct the image covariance
matrix. This improves the computational effciency. Moreover, the projection of
sample on each principal orthogonal vector is a vector. A drawback of 2DPCA is
that it needs more coeffcients than PCA for image representation and costs more
time to calculate distance in classifcation phase. In order to address this problem,
Ye (2004) proposed a new algorithm called generalized low rank approximations
of matrices (GLRAM) to reduce the computational cost. Then some researchers
proposed a non-iterative algorithm for GLRAM (Liang & Shi, 2005; Liang, Zhang,
& Shi, 2007). Moreover, they reveal the optimal property of GLRAM and show
that the reconstruction errors of GLRAM are not smaller than those of PCA when
considering the same dimension. Likewise, their method is proved to have much
less computational time than the traditional singular value decomposition (SVD)
technique. In addition, researchers also developed a number of variants of 2DPCA
(Xu et al., 2005; Nhat & Lee, 2005; Xu, Jin, Jiang, & Guo, 2006; Hou, Gao, Pan, &
Zhang, 2006; Vasilescu & Terzopoulos, 2002; Wang & Ahuja, 2004; Zuo Wang, &
Zhang, 2005a, 2005b). In fact, the methods mentioned above belong to the frame-
work proposed by Lathauwer and his partners (Lathauwer, 1997; Lathauwer, Moor,
& Vandewalle, 2000a). In 2000, a multilinear generalization of the singular value
decomposition was further proposed (Lathauwer, Moor, & Vandewalle, 2000b).
Moreover, they also analyzed some properties of the matrix and the higher-order
tensor decompositions. Yu and Bennamoun (2006) also proposed nD-PCA algo-
rithm which exploits higher-order singular value decomposition. All the methods
contribute to the development of the tensor PCA.
is prohibited.
8.2 bAsIc Algor Ithms
8.2.1 classical PcA
Assume that x
1
,...x
M
are M data points in R
N
. Without loss of generality, we assume
that the data are centered, which shows that the mean of all data is zero. In fact,
they can be easily achieved by performing a translation of data. Generally speak-
ing, PCA extracts the desired number of principal components for data. In other
words, PCA is to fnd a matrix W that maps these M points to another set of points
y
1
,...,y
M
in R
d
, where d < N.
From the viewpoint of data reconstruction, PCA can be considered to fnd the
matrix W = (w
1
,...,w
d
) by minimizing the average L
2
reconstruction errors for all
data, denoted by
2
1
1
( ) ( )
M
T
i i i i
i
x x w x x
M
=
= -

. (8.1)
The transformation matrix W can be obtained in the following way. Let X be an
N M matrix whose columns consist of x
1
,...x
M
. Then the matrix W is obtained by
the eigen-decomposition of /
T
C XX M = whose columns consist of the eigenvectors
corresponding to the frst m largest eigenvalues of C. It is not diffcult to verify that
the ith eigenvalue is the variance of the data projection on w
i
. In addition, another
good property of PCA is that it decorates the data.
8.2.2 t wo-dimensional PcA
Since it involves image matrices in 2DPCA, it is useful to introduce image projections
before we begin. Let {A
1
,...,A
M
} be M image matrices, where A
i
is an m n matrix.
Then the image A is projected onto a n-dimensional vector v. That is, the following
transformation is adopted: Y
i
= A
i
v. It is obvious that the projection vector Y
i
is an
m-dimensional vector. In order to measure the discriminatory power of vector v,
the total scatter of the projected vectors is introduced. In Yang, Zhang, Frangi and
Yang (2004), the total scatter of the projected samples is defned by the trace of the
covariance matrix of the projected feature vectors, which is denoted as follows:
J(X) = trace(S
x
), (8.2)
where S
x
denotes the covariance matrix of the projected feature vectors of training
samples.
is prohibited.
Using the defnition of the image covariance matrix, one can obtain
1
1
( ) ( ( ) ( ))
M
T T
i i
i
J X v A A A A v
M
=
= - -
= v
T
G
t
v, (8.3)
where A denotes the average image of all training samples.
In general, the optimal projected vector v is the eigenvector corresponding to
the largest eigenvalue of G
t
. In some cases, only a projected vector is not enough.
It is necessary to obtain a set of the projected vectors v
1
,...,v
d
which can maximize
the criterion function J(X). In such a case, the optimal projected vectors v
1
,...,v
d
are
the orthogonal eigenvectors of G
t
corresponding to the frst d largest eigenvalues.
8.2.3 generalized Low Rank Approximations of Matrices
Consider a p q-dimensional space L R, where denotes the tensor product, L
is spanned by {u
1
,...,u
p
}, and R is spanned by {v
1
,...,v
q
}. Then the following matrices
are defned: L = [u
1
,...,u
p
] and R = [v
1
,...,v
q
]. The GLRAM method is to fnd L and
R so that LD
i
R
T
is a good approximation of A
i
(i = 1,...,s), where
p q
i
D

. In other
words, the GLRAM method is to solve the following optimization problem:
2
:
1
:
1
min
r p T
p
c q T
q
p q
i
M
T
i i
F
L L L I
i
R R R I
D
A LD R
M
=
=
=
. (8.4)
Compared with the optimization problem in Ye (2004), we add a constant in Eq.
(8.4). However, this does not affect the following discussion.
It has theoretically proved that the minimization problem of Eq. (8.4) is equivalent
to the following maximization problem:
2
:
1
:
1
max
r p T
p
c q T
q
M
T
i
F
L L L I
i
R R R I
L AR
M
=
=
=
. (8.5)
Ye noticed that there is no closed solution of Eq. (8.5). To this end, the following
theorem which is very useful for obtaining the solution of Eq. (8.5) is proposed.
Theorem 8.1 Let L, R and
1
{ }
M
i i
D
=
be the optimal solution to the minimization
problem in Eq. (8.4). Then
1. For a given R,L consists of the p eigenvectors of the matrix
is prohibited.
1
1
M
T T
L i i
i
D ARR A
M
=
=

corresponding to the largest p eigenvalues.
2. For a given L,R consists of the q eigenvectors of the matrix
1
1
M
T T
R i i
i
D A LL A
M
=
=

corresponding to the largest q eigenvalues.
By virtue of the idea of Theorem 8.1, an iterative procedure is proposed for
obtaining L and R. A much more detailed discussion for the iterative procedure
can be found in Ye (2004).
8.2.4 non-Iterative glr Am
Before non-iterative GLRAM is developed, it is necessary to introduce lemma 8.2
and its corollary.
Lemma8.2 (Schott, 1997): Let B be an N N symmetric matrix and H be an N
h matrix which satisfes H
T
H = I
h
, where I
h
is a h h identity matrix. Then, for i
= 1,...,h, we have
( ) ( )
T
N h i i i
H BH B
- +

, (8.6)
where
i
(B) denotes the ith largest eigenvalue of the matrix B.
From Lemma 8.2, it is obvious that the following corollary can be obtained.
Corollary8.3 Let w
i
be the eigenvector corresponding to the ith largest eigenvalue
i
of B and H be an N h matrix which satisfes H
T
H = I
h
. Then
1 1
( )
T
N h m h
trace H BH
- +
+ + + + , (8.7)
the second equality holds if H = WQ, where W = (w
1
,...,w
h
) and Q is any h h or-
thogonal matrix.
From corollary 8.3, it is not diffcult to obtain the following theorem.
Theorem8.4 Let w
i
be the eigenvector corresponding to the ith largest eigenvalue
i
of C = XX
T
/ M (i = 1,...,mn). Then
is prohibited.
1 r c p q m n - +
+ +
2
1
1
M
T
i
F
i
L AR
M
=

1 p q
+ +
. (8.8)
Proof:
2
1
M
T
i
F
i
L AR
=

2
2
1
( )
M
T
i
i
vec L AR
=
=
, (8.9)
where vec denotes the vec operator which can convert the matrix to a vector by
stacking the columns of matrix.
Note that vec(L
T
A
i
R) = (R
T
L
T
)vecA
i
. For the sake of notational simplicity, let
W
T
= (R
T
L
T
). In such a case, Eq. (8.9) is further represented as
2
2
1
( )
M
T
i
i
vec L AR
=

1
( ( ) ( ) )
M
T T
i i
i
tr W vec A vec A W
=
( )
T T
tr W XX W = .
That is
2
1
1
M
T
i
F
i
L AR
M
=

1
( )
T T
tr W XX W
M
= .
Note that W
T
W= (R
T
L
T
)(R L) = (R
T
R) (L
T
L) = I
q
I
p
= I
pq
. In such a case,
applying Corollary 8.3, we can obtain
1 m n p q m n - +
+ +
2
1
1
M
T
i
F
i
L AR
M
=

1 p q
+ +
. (8.10)
This completes the proof of the theorem.
Theorem 8.4 demonstrates the bound for the objective function in Eq. (8.5).
However, it does not provide the solution for GLRAM. It is obvious that the objec-
tive function obtains the optimal value when W consists of the eigenvectors of C
corresponding to the frst pq largest eigenvalues. To our knowledge, it is easy to
compute the Kronecker product of two matrices. It is, however, very diffcult or
impossible to decompose a large matrix into the form of the Kronecker product of
two matrices. In addition, it is straightforward to verify that the second equality
in Eq. (8.5) holds if A
i
(i = 1,...,M) are m 1 matrices, namely, vectors. It is clear
is prohibited.
that this is the form of classical PCA. Therefore, in some sense, PCA is a special
case of GLRAM when the objective function achieves the optimal value. From
the viewpoint of data reconstruction, we can see that the reconstruction errors
obtained by GLRAM are not smaller than those obtained by PCA when the same
dimensionality which has reduced is considered.
To further develop the theory, the following matrix is defned (Liang et al.,
2007):
G
s1
=
1
1
( ) ( )
M
T
i i
i
A A
M
=
, G
s2
=
1
1
( )( )
M
T
i i
i
A A
M
=
. (8.11)
Assume that F
1
consists of the eigenvectors of G
s1
corresponding to the frst q
largest eigenvalues and F
2
consists of the eigenvectors of G
s2
frst p largest eigenvalues. Let
1 1 1
1
1
M
T T
L i i
i
D AF F A
M
=
=

,
1 2 2
1
1
M
T T
R i i
i
D A F F A
M
=
=

. (8.12)
Assume that K
1
consists of the eigenvectors of D
L1
corresponding to the frst
largest eigenvalues and K
2
consists of the eigenvectors of D
R1
frst q largest eigenvalues.
Applying Corollary 8.3 in such a case, it is not diffcult to obtain the following
theorem.
Theorem8.5 Letd1be the sum ofthe frst p largest eigenvaules of D
L1
andd2be
the sum ofthe frst q largest eigenvaules of D
R1
. In such a case,the maximal value
of Eq. (8.5) is equal to max{d1, d2}.
Proof: (a) Note that L and R maximize
2
1
1
M
T
i
F
i
L AR
M
=
. (8.13)
Equation (8.13) is further represented as
2
1 1
1 1
( )
M M
T T T T
i i i
F
i i
L AR trace L ARR A L
M M
= =
=

is prohibited.
1
1
( ( ) )
M
T T T
i i
i
trace L ARR A L
M
=
=

= trace(L
T
D
L
L).
Applying Corollary 8.3, we have
trace(L
T
D
L
L) trace(D
L
)
p
. (8.14)
Since trace(D
L
) =
1
1
( ( ))
M
T T
i i
i
trace ARR A
M
=
= trace(R
T
G
s1
R) , we have
trace(R
T
G
s1
R) trace(G
s1
)
q
. (8.15)
From Eq. (8.14) and Eq. (8.15), we obtain
trace(L
T
G
s1
L) trace(G
s1
)
q
.
From Eq. (8.15) and corollary 8.3, it is not diffcult to obtain R, denoted as,
2
1 q q
R FQ

=
, (8.16)
where
2
q q
Q

is any orthogonal matrix.
Substituting Eq. (8.16) into D
L
, we obtain D
L1
in Eq. (3.12).
From Eq. (8.14) and Corollary 8.3, it is not diffcult to obtain L, denoted as,
1
1 p p
L K Q

= , (8.17)
where
1
p p
Q

Furthermore, it is straightforward to verify that the maximal value of Eq. (8.5)
is equal to d1
(b)
2
1 1
1 1
( )
M M
T T T T
i i i
F
i i
L AR trace R A LL A R
M M
= =
=

trace(R
T
D
R
R)
Applying Corollary 8.3, we have
( ) ( )
T
R R q
trace R D R trace D . (8.18)
is prohibited.
Since trace(D
R
) =
1
1
( )
M
T T
i i
i
trace L A A L
M
=
= trace(L
T
G
s2
L), we have
trace(L
T
G
s2
L) trace(G
s2
)
p
. (8.19)
From Eq. (8.17) and Eq. (8.18), we obtain trace(R
T
D
L
R) trace(G
s2
)
p
.
From Eq. (8.18) and Corollary 8.3, it is not diffcult to obtain L, denoted as,
1
2 p p
L F Q

=
, (8.20)
Substituting Eq. (8.19) into D
R
, we obtain D
R1
in Eq. (8.12). Similarly, we can
obtain R, denoted as
2
2 q q
R K Q

=
, (8.21)
where
2
q q
Q

In such a case it is straightforward to verify that the maximal value of Eq. (8.5)
is equal to d2.
From (a) and (b), this completes the proof of the theorem.
From the proof of Theorem 8.4, it is not diffcult to fnd the algorithm for GL-
RAM. Finally, the method is summarized as follows.
The non-iterative GLRAM algorithm
Step1: Assume that p,q are given. Compute the matrices G
s1
and G
s2
.
Step2: Compute the eigenvectors of G
s1
and G
s2
,let
2
2 q q
R K Q

= and
1
2 p p
L F Q

=
Step 3: Compute the eigenvectors of D
L1
and D
R1
, and obtain
1
1 p p
L K Q

= corre-
sponding to R in step 2 and
2
1 q q
R FQ

= corresponding to Lin step 2, and compute
d1 and d2.
Step4: Choose R, L corresponding to max{d1, d2}, and compute D
i
= L
T
A
i
R.
In addition, for the non-iterative algorithms, the following remarks are given.
1. We can see from the proof of Theorem 8.5 that two different solutions are
obtained. Note that although
T
i i
A A and
T
i i
A A have the same nonzero eigenval-
ues, G
s1
and G
s2
may have different nonzero eigenvalues. But traceG
s1
and
traceG
s2
are equal. Therefore, the objective values corresponding to differ-
ent solutions are different in general cases. In other words, d1 and d2 are not
in general equal. To this end, we choose the bigger value in order to obtain
smaller reconstruction errors. Meanwhile, we notice that d1 and d2 are equal
if A
i
(i = 1,...,s) are symmetric square matrices.
is prohibited.
2. We can also notice that the reconstruction errors obtained by the non-iterative
algorithm for GLRAM are not bigger than those obtained by Yes iterative
algorithm (2004) because Yes solution may be locally optimal. Likewise, the
reconstruction errors obtained by non-iterative algorithm are not smaller than
those obtained by PCA.
3. We can see that L and R are not unique. However, the solutions subject to an
orthogonal transformation. In general, for the sake of simplicity,
2
q q
Q

and
1
p p
Q

are chosen as identity matrices in real applications.
4. It is straightforward to verify that min{p,q} affects the performance of gener-
alized low rank approximations of matrices. In general, the bigger min{p,q}
is, the better the approximation performance is. However, it increases the
computational cost. Hence, we must make a compromise between the recon-
struction errors and the computational cost. Meanwhile, we also notice that
a simple strategy for choosing {p,q} is provided. That is, assume that the
ratio of lost information in the second step of the algorithm is smaller than
the threshold value. Then we choose {p,q}. In a similar way, we can choose
another corresponding dimensionality by setting another threshold value. In
such a case, we can choose the solution with smaller dimensions.
5. It is of interest to note that the above algorithm is related to 2DPCA. Assume
that E(A) = 0
rc
, L = I
rr
, and R = V. It is clear that the method for generalized
low rank approximations of matrices is equivalent to 2DPCA in such a case. In
other words, the method for generalized low rank approximations of matrices
is a generalization of 2DPCA. In addition, we can know that the distributions
of eigenvalues have a similar characteristic of those of PCA (see Theorem
8.4).
6. Compared with Yes iterative algorithm (2004), the non-iterative algorithm for
GLRAM has smaller computational complexity. However, it is not superior
to 2DPCA in terms of the computational effciency because the non-iterative
algorithm for GLRAM needs to run several times (each time corresponds to a
2DPCA algorithm). It is obvious that the non-iterative algorithm for GLRAM
has lower computational cost than PCA since the non-iterative algorithm for
GLRAM needs to solve a small eigenvalue problem.
8.2.5 higher-Order PCA
Lathauwer et al. (2000a) developed a higher-order tensor based the singular value
decomposition, which provides the basis for developing some variants of higher-
order PCA. Based on their theories, it is easy to use the ideas of tensors to represent
higher dimensional datasets.
is prohibited.
For the sake of clarity, we restate this Nth-order SVD as shown in Chapter
VII.
Theorem8.6(Lathauwer et al., 2000a) Each complex (I
1
I
2
... I
N
)-tensor A can
be written as the product
A = S
1
U
(1)

2
U
(2)
...
N
U
(N)
, in which
1.
( ) ( ) ( ) ( )
1 2
(
n
n n n n
I
U U U U = ) is a unitary (I
n
I
n
)-matrix,
2. S is a complex (I
1
I
2
... I
N
)-tensor of which the subtensor
n
i
S
=
, obtained by
fxing the nth index to , have the properties of
i. all-orthogonality: two subtensors
n
i
S
=
and
n
i
S
=
are orthogonal for all possible
of n, subject to :
<
n
i
S
=
,
n
i
S
=
>=0 when ,
ii. ordering:
1
n
i
S
=

2
n
i
S
=
0
n
n
I
i
S
=
for all possible values of n.
The Frobenious-norms
n
i
i
S
=
are n-mode singular values of A and the vector
( ) n
i
U
is an ith n-mode singular vector.
Note that the mode-n multiplication in theorem 8.6 is frst implemented by un-
folding the tensor A along the given mode-n to generate its model-n matrix form
A
(n)
, and then performing the matrix multiplication. After that, the resulting matrix
B
(n)
is folded back to the tensor form. Lathauwer et al. (2000a) pointed out that it
is not impossible to reduce the higher-order tensors to a pseudodiagnonal form by
means of the orthogonal transformation.
Yu and Bennamoun (2006) constructed a different tensor as follows:
1
(( ), , ( ))
N
D X X X X = - -
where
1 N
I I
i
X R

. It is obvious that D is a tenor.
In such a case, applying Theorem 8.6 to D will produce a mode-n singular vector
contained in
( )
, 1, , .
n
U n N = After that, one can determine the desired principal
orthogonal vectors for each mode of the tensor D in terms of the mode-n singular
values. Then the projection of sample X onto the mode-n principal vector
( ) n
k
U is
expressed as
( )
( ) ( )
n T
n n k
Y X X U = - .
It is straightforward to verify that the projection Y
n
is still an N-order tensor. Pro-
jecting X onto a principal vector
( ) n
i
u of
( ) n
k
U will produce
( )
,
( ) ( ) ,
n T
n i n k
Y X X u i k = -
.
is prohibited.
The projection Y
n,i
is a (N-1)-order tensor. The fnal construction of X can be obtained
by
( ) n
n n k
X X Y U = +
. For classifcation, the Forbenious-norms between two mode-n

principal component tensors,
i
n
Y
and
j
n
Y
(i j ) are adopted as follows.
( , )
i j i j
n n n n
F
d Y Y Y Y = - .
8.3 APPl IcAt Ions t o bI ometr Ic ver If IcAt Ion
In this section, experiments on face images and handwritten characters are made
to test PCA and its several variants. From the discussion in section 8.2, we can see
that two important parameters p and q are involved in GLRAM. For the sake of
simplicity, p and q are set to the common value d in the following experiments. As
pointed out in Ye (2004), this choice is a good strategy for obtaining good classif-
cation performance. In addition, when the classifcation performance is involved,
it is necessary to use a classifer for evaluation. To this end, the nearest neighbor
classifer is adopted due to its simplicity.
8.3.1 experiments on the orl f ace database
Each face image of the ORL database (please refer to Section 3.2.3 ) is downsampled
to 28 23 pixels to reduce the computational complexity. First, we perform experi-
ments to check the distributions of eigenvalues of the non-iterative algorithm for
GLRAM. Figure 8.1(a) denotes the distributions of eigenvalues of Eq. (8.11) and
Figure 8.1 (b) denotes the distributions of eigenvalues of Eq. (8.12) when F
1
chooses
different dimensions. As observed in Figure 8.1, the magnitude of eigenvalues
decreases quickly. The frst several eigenvalues contain most information (energy)
of the original image. As a result, it is very reasonable to choose the frst several
eigenvectors to approximate the original image so that the dimensions of images can
be drastically reduced, which is very useful for image retrieval and recognition.
In the following, experiments are made to obtain the reconstructed images.
Figure 8.2 shows the reconstructed images of four methods, namely PCA, 2DPCA,
GLRAM, and the non-iterative algorithm for GLRAM (NGLRAM), by adding the
dimensions of features. It should be noted that the dimension of 2DPCA is 28 d,
while the dimensions of other three methods are d d. As can be seen from Figure
8.2, the visual quality of the reconstructed images is improved with the increase
of features. It is obvious that the reconstructed images obtained by the non-itera-
tive algorithm for GLRAM have slightly better visual quality than those obtained
by GLRAM, while they are not superior to those obtained by PCA. When the frst
is prohibited.
Figure 8.1 (a). The distributions of eigenvalues in q, (b) the distributions of eigen-
values in p.
(a)
(b)
several eigenvectors are used to reconstruct the images, PCA shows the contours
of images, whereas other three methods cannot show the contours of images.
This set of experiments was used to compare non-iterative GLRAM with
2DPCA, PCA and GLRAM in terms of the reconstruction errors and the recogni-
is prohibited.
tion performance. We use a tenfold cross-validation to evaluate the classifcation
performance. Note that the reconstruction errors are defned as a square root of
Eq. (8.5). Figure 8.2 (a) denotes the reconstruction errors obtained by four methods
with the change of features and Figure 8.3 (b) denotes the recognition performance
obtained by four methods with the change of features. As can be seen from Fig-
ure 8.3 (a), the reconstruction errors decrease with the increase of features. The
PCA method obtains the smallest reconstruction errors in all methods. It is clear
that the non-iterative algorithm for GLRAM is superior to 2DPCA or GLRAM in
terms of the reconstruction errors. When the dimension is small, GLRAM obtains
smaller reconstruction errors than 2DPCA. However, with the increase of features,
the reconstruction errors of 2DPCA are smaller than those of GLRAM. Here, we
also explain this reason. When the dimension m = p q is small, p is a constant in
2DPCA. Therefore, q is smaller, and thus results in large information lost. How-
ever, for GLRAM and our method, we make p = q. In such a case, we can obtain
bigger p and q. Consequently, the information is lost in two stages in GLRAM. As
discussed in Ye (2004), GLRAM may obtain locally optimal solutions, which helps
us to explain this phenomenon. As observed in Figure 8.3 (b), in most cases, the
recognition performance increases with the increase of features. When the number
of features is small, the 2DPCA method is not superior to other three methods in
terms of the recognition performance. However, with the increase of features, these
four methods can obtain similar results, which show the matrix-based methods have
potential applications for image recognition.
In addition, we also conduct the experiments to compare the classifcation per-
formance of non-iterative GLRAM with those of 2DPCA, PCA and GLRAM, under
Figure 8.2. The reconstruction images of four methods (d=1:23)
(a) GLRAM
(b) NGLRAM
(c) PCA
(d) 2DPCA
is prohibited.
Figure 8.3 (a) reconstruction errors obtained by four methods with the change of
features. (b) recognition performance obtained by four methods with the change
of features.
(a)
(b)
the conditions where the number of training samples is varied. To reduce variation,
the experimental results reported in this experiment are obtained by 20 runs. Figure
8.4 denotes the recognition performance of four methods with the change of the
is prohibited.
number of training samples. Here we set the number of features to 28 7 (14 14).
From Figure 8.4, we can see that PCA obtains the best classifcation performance
in all methods, which shows the matrix-based methods are not always superior to
PCA. However, we know that the matrix-based methods are superior to PCA in
terms of the computational effciency in dealing with large image matrices (Yang et
al., 2004). It is also observed that the non-iterative algorithm for GLRAM obtains
slightly better classifcation performance than 2DPCA or GLRAM.
8.3.2 Experiments on handwritten Numeral Character Database
The handwritten numeral characters are obtained from the UCI data set repository.
This data set consists of 5620 handwritten numeral characters. The original im-
age of each character has the size of 32 32 pixels. In order to reduce the size of
the image, we obtain the size of 16 16 pixels where each pixel is obtained from
the average of the block of 2 2 pixels in the original images. In other words, the
number of features of each character image is 256. In our experiments, we only
choose 100 samples from each class to constitute training samples and others are
used for testing. As a result, the number of training samples is 1000 and the number
of testing samples is 4620. To enhance the accuracy of performance evaluation,
the classifcation performance reported in our experiments is averaged over ten
Figure 8.4. Recognition performance of four methods with the change of the number
of training samples
is prohibited.
times. In other words, 10 different training and testing sets are used for evaluating
the performance.
As discussed in the ORL experiments in Section 8.3.1, the frst experiment is
made to show the distributions of eigenvalues. Figure 8.5 (a) denotes the distribu-
tions of eigenvalues of Eq. (8.11) and Figure 8.5 (b) denotes the distributions of
eigenvalues of Eq. (8.12) when F
1
chooses different dimensions. As can be seen from
Figure 8.5, the magnitudes of eigenvalues also decrease. It shows that extracting
the frst several eigenvectors to preserve the majority of information of the original
images is enough.
Then we conduct experiments to compare non-iterative GLRAM with 2DPCA,
PCA and GLRAM in terms of the reconstructed images. Figure 8.6 denotes the
reconstruction images of handwritten numeral character 2. It is obvious that the
images become clearer with the increase of features in most cases. We can see that
the reconstruction images obtained by the non-iterative algorithm for GLRAM
have slightly better visual quality than those obtained by GLRAM. As discussed
in Section 8.3.1, PCA achieves the best reconstructed images in all methods.
In addition, experiments are performed to compare the non-iterative GLRAM
with 2DPCA, PCA and GLRAM in terms of the reconstruction errors and the clas-
sifcation performance. Figure 8.7 (a) denotes the reconstruction errors obtained by
four methods with the change of features and Figure 8.7 (b) denotes the recognition
performance obtained by four methods with the change of features. From Figure
8.7 (a), it is clear that the reconstructed errors of four methods decrease with the
increase of features. The PCA method obtains the smallest reconstruction errors
in all methods. Like experiments in Section 8.3.1, GLRAM obtains smaller recon-
struction errors than 2DPCA when the number of features is small. However, with
the increase of features, the reconstruction errors between 2DPCA and GLRAM
make no remarkable difference. It is different from experimental results in Section
8.3.1. Probably, a relatively large number of samples are used as training samples
in this experiment. But it is clear that the non-iterative algorithm for GLRAM is
superior to 2DPCA or GLRAM in terms of the reconstruction errors. Note that the
reconstruction errors obtained by the non-iterative algorithm for GLRAM approach
those obtained by PCA in this experiment. This may come from the fact that the
rank of the covariance matrix is relatively big compared with the dimensions of
samples. As observed in Figure 8.7 (b), the classifcation performance is improved
with the increase of features in most cases. However, 2DPCA still obtains higher
recognition rates even when the number of features is 2. But there is no remarkable
difference in the classifcation performance for four methods when the number
of features is big. Therefore, since matrix-based methods are more effcient than
PCA in terms of the computational time, matrix-based methods such as 2DPCA
and GLRAM can be used to recognize images even when a number of training
samples are available.
is prohibited.
Figure 8.5 (a) The distributions of eigenvalues in q, (b) the distributions of eigen-
values in p.
(a)
(b)
is prohibited.
Figure 8.6. The reconstructions images of four methods (d=1:16)

(a) GLRAM
(b) NGLRAM

(c) PCA
(d) 2DPCA
8.4 summAr y And dI scuss Ion
Tensor PCA is an effective method for data reconstruction and recognition. How-
ever, tensor PCA adopts an iterative procedure to extract features. In this chapter,
we discuss the properties of tensor PCA. It is of interest to note that the optimal
value of tensor PCA is closely related to classical PCA. We theoretically show the
reconstruction errors obtained by GLRAM (a special of tensor PCA) are not smaller
than those obtained by PCA when the same number of dimensionality is considered.
We also show that the reconstruction errors obtained by non-iterative GLRAM
method are still not bigger than those obtained by Yes iterative algorithm (2004)
when considering the same dimensionality. In the meantime, a simple strategy for
choosing dimensions is provided.
A possible explanation that PCA has smaller reconstruction errors than matrix-
based methods even high-order methods is that the covariance matrix of classical
PCA is positive semi-defnite, especially for the small sample-size problem, which
makes the data information tight, while the image (tensor) covariance matrix in the
matrix (tensor)-based methods is positive defnite in general cases, which results
in information lost as long as the dimension is reduced.
Experimental results on face images and handwritten numerical characters
demonstrate that tensor PCA is still competitive with other existing methods such
as classical PCA in terms of the classifcation performance or the reconstructed
errors.
However, there are still several aspects needed to be pointed out. On the one
hand, the computational complexity is still a challenging problem in the case of
tensor PCA due to the high dimension arrays. Consequently, how to improve the
is prohibited.
(a)
(b)
Figure 8.7 (a). reconstruction errors obtained by four methods with the change of
features. (b) Recognition performance obtained by four methods with the change
of features.
computational speed needs to be further considered. One the other hand, whether
there exists the condition such that the reconstruction errors of tensor PCA are
equal to those of PCA is unclear up till now and whether the solution obtained by
the iterative algorithm is optimal needs to be further investigated. But we expect
is prohibited.
that the insights in this chapter will pave the way for building close relationship
between classical PCA and tensor PCA.
r eferences
Hou, J., Gao, Q. X., Pan, Q., & Zhang, H. C. (2006). Essence of 2DPCA and modi-
fcation method for face recognition. Proceedings of the International Conference
on Machine learning and Cybernetic (pp. 3351-3353).
Lathauwer, L. D. (1997). Signal processing based on multilinear algebra. Ph.D
thesis, K.U. Leuven, Belgium.
Lathauwer, L., Moor, B., & Vandewalle, J. (2000a). A multilinear singular value
decomposition. J. Matrix Anal. Appl., 21(4), 1253-1278.
Lathauwer, L., Moor, B., & Vandewalle, J. (2000b). On the best of rank-1 and rank-
(R
1
,R
2,
R
N
) approximation of high-order tensors. J. Matrix Anal. Appl., 21(4),
1324-1342.
Liang, Z. Z., & Shi, P. F. (2005). An analytical method for generalized low rank
approximation matrices. Pattern Recognition, 38(11), 2213-2216.
Liang, Z. Z., Zhang, D., & Shi, P. F. (2007). The theoretical analysis of GRLDA
and its applications. Pattern Recognition, 40(3), 1032-1041.
Nhat, V. D., & Lee, S. Y. (2005). Two-dimensional weighted PCA algorithm for face
recognition. Proceedings of the IEEE Symposium on Computational intelligence
in Robotics and Automation (pp. 219-223).
Penev, P. S., & Sirovich, L. (2000). The global dimensionality of face space. Proc.
fourth IEEE Intl Conference Automatic face and Gesture Recognition (pp. 264-
270).
Schott, J. R. (1997). Matrix analysis for statistics. New York: Wiley InterScience
publication.
Turk, M., & Pentland, A. (1991). Eigenface to recognition. J. cognitive Neurosience,
3(1), 71-86.
Vasilescu, M. A., & Terzopoulos, D. (2002). Multilinear image analysis for facial
recognition. Proceedings of the 16th International Conference on Pattern Recogni-
tion (ICPR) (pp. 511-514).
is prohibited.
Wang, H. C., & Ahuja, N. (2004). Compact representation of multidimensional
data using tensor rank-one decomposition. Proceedings of the 17th International
Conference on Pattern Recognition (ICPR) (pp. 44-47).
Xu, A. B., Jin, X., Jiang, Y. G., & Guo, P. (2006). Complete two-dimensional PCA
for face recognition. Proceedings of the 18th International Conference on Pattern
Recognition (ICPR) (pp. 481-484).
Xu, D., Yan, S. C., Zhang, L., Li, M. J., Ma, W. Y., & Liu, Z. K., et al. (2005).
Parallel image matrix compression for face recognition. Proceedings of the 11th
International Conference on Multimedia Modeling (pp. 232-238).
Yang, J., Zhang, D., Frangi, A. F., & Yang, J-Y. (2004). Two-dimensional PCA: a
new approach to appearance-based face representation and recognition. IEEE Trans.
Pattern Anal. Machine Intell., 26(1), 131-137.
Ye,J. P. (2004). Generalized low rank approximations of matrices. International
Conference on Machine Learning Conference Proceedings (pp. 887894).
Yu, H., & Bennamoun, M. (2006). 1D-PCA, 2D-PCA to nD-PCA. Proceedings of
the 18th International Conference on Pattern Recognition (ICPR) (pp. 181-184).
Zuo, W. M., Wang, K. Q., & Zhang, D. (2005a). Assembled matrix distance metric
for 2DPCA-based face and palmprint recognition. Proceedings of the International
Conference on Machine Learning and Cybernetics (pp. 4870-4875).
Zuo, W. M., Wang, K. Q., & Zhang, D. (2005b). Bi-directional PCA with assembled
matrix distance metric. Proceedings of the International Conference on Image
Processing (pp. 958-961).
is prohibited.
Chapter IX
Tensor Linear Discriminant
Analysis
Abstr Act
Linear discriminant analysis is a very effective and important method for feature
extraction. In general, image matrices are often transformed into vectors prior to
feature extraction, which results in the curse of dimensionality when the dimensions
of matrices are huge. In this chapter, classical LDA and its several variants are
introduced. In some sense, the variants of LDA can avoid the singularity problem
and achieve computational effciency. Experimental results on biometric data show
the usefulness of LDA and its variants in some cases.
9.1 Introduct Ion
Linear discriminant analysis is a popular technique for feature extraction, which
has been successfully applied in many felds such as face recognition and character
recognition. Linear discriminant analysis seeks to fnd the direction which maxi-
mizes between-class scatter and minimizes the within-class scatter. Based on linear
discriminant analysis, Foley and Sammon (1975) proposed optimal discriminant
vectors for two-class problems. Duchene and Leclercq (1988) further presented a
set of discriminant vectors to solve multi-class problems. Although Foley-Sammon
optimal discriminant vectors (FSODV) are orthogonal and perform well in some
cases, the features which are obtained by optimal orthogonal discriminant vectors
are statistically correlated. To avoid this problem, Jin, Yang, Hu, and Luo (2001)
Tensor Linear Discriminant Analysis 173
is prohibited.
proposed a new set of uncorrelated discriminant vectors(UDV) which is proved
to be more powerful than that of optimal orthogonal discriminant vectors in some
cases. Then Jing, Zhang, and Jin (2003) further stated improvements on uncorre-
lated optimal discriminant vectors. Subsequently, Xu, Yang, and Jin (2003) studied
the relationship between the Fisher criterion values of FSODV and UDV. Then
Xu, Yang, and Jin (2004) developed a new model for Fisher discriminant analysis,
which applies the maximal Fisher criterion and the minimal statistical correlation
between features. Since the methods mentioned above are based on vectors rather
than matrices, these methods face the computational diffculty when the dimension
of data is too huge. To overcome this problem, Liu, Cheng, and Yang (1993) frstly
proposed a novel linear projection method, which performs linear discriminant
analysis in terms of image matrices. However, feature vectors using Lius method
could be statistically correlated. In order to effectively deal with this problem, Yang,
Yang, Frangi, and Zhang (2003) proposed a set of two-dimensional (2D) projection
vectors which satisfy conjugate orthogonal constraints. Most importantly, feature
vectors obtained by Yangs method are statically uncorrelated. Then Liang, Shi, and
Zhang (2006) proposed a new technique for 2D Fisher discriminant analysis. In their
algorithm, the Fisher criterion function is directly constructed in terms of image
matrices. Then they utilize the Fisher criterion and statistical correlation between
features to construct an objective function. Then discriminant vectors are obtained
in terms of the objection function. At the same time, they theoretically analyze that
the proposed algorithm is equivalent to uncorrelated two-dimensional discriminant
analysis in some condition. In Xiong, Swam and Ahmad (2005), one-sided 2DLDA
is developed for classifcation tasks. Ye, Janardan, Park, and Park (2004) further
developed generalized 2DLDA, which can overcome the singularity problem and
achieves the computational effciency. In Liang (2006; Yan et al., 2005), a multilinear
generalization of linear discriminant analysis is discussed and an iterative algorithm
is developed for solving multilinear linear discriminant analysis. Meanwhile, a
non-iterative algorithm is also proposed in Liang (2006). In addition, multilinear
LDA provides a unifed framework for classical LDA and 2DLDA.
9.2 bAsIc Algor Ithms
Notations: Let {A
1
,...,A
M
} denote a set of images, A
i
R
mn
. Each image belongs to
exactly one of c object class {l
1
,...,l
c
}. The number of images in class l
i
is denoted by
n
i
and
i
n M =
. Let x
i
= vec(A
i
),where vec denotes the vector operator which can
convert the matrix by stacking the column of the matrix. x
i
R
N
, N = m n.
is prohibited.
9.2.1 Classical Linear Discriminant Analysis
In discriminant analysis, the within-class and between-class scatter matrices are
defned as follows:
1
( )( ) ,
c
T T
b i i i b b
i
S n m m m m H H
=
= - - =
(9.1)
1
( )( ) ,
i
c
T T
w i i w w
i x l
S x m x m H H
=
= - - =
(9.2)
where the precursors H
b
and H
w
of the between-class and within-class scatter ma-
trices in (9.1) and (9.2) are
1 1
[ ( ), , ( )]
b c c
H n m m n m m = - - , (9.3)
1 1 1
[ , , )]
T T
w c c c
H X me X m e = - - , (9.4)
(1, ,1)
i
n T
i
e =
, X
i
is the data matrix for class l
i
, m
i
is the centroid of the ith class
and m is the global centroid of the data set.
Classical LDA is to compute a linear transformation G R
Nd
that maps x
i
in the
N-dimensional space to a vector
i
x in the d-dimensional space by
T
i i
x G x = . Applying
the linear transformation matrix G, one can obtain the within-class and between-
class distance in the projected space. Then one can maximize the between-class
distance and minimize the within-class distance in the projection space. Often, the
optimization criterion in classical LDA can be defned as follows:
( ) max ( )
T
b
T
w
G S G
J G trace
G S G
= (9.5)
The optimization problem can be solved by applying the eigen-decomposition on
the matrix
1
w b
S S
-
if S
w
is nonsingular. In general, there are at most c 1 eigenvectors
corresponding to nonzero eigenvalues since the rank of the matrix S
b
is not bigger
than c 1. In the case of singularity, some effective methods (Ye, 2005; Ye & Li,
2005) have been developed in recent years.
9.2.2 t wo-dimensional ld A
Without loss of generality, let the image A be an m n matrix. Then the image A
is projected onto an n-dimensional vector X . That is, the following transformation
is adopted:
is prohibited.
Y = AX. (9.6)
It is obvious that the projection vector Y is an m-dimensional vector. In this
case, the image A is transformed into an m-dimensional vector Y and each image
corresponds to a vector.
Let A
ij
denote the ith image in the jth class. Then we project the image onto X
and obtain the following form:
ij ij
Y A X = , 1, 2, , , 1, 2, ,
j
i n j c = = . (9.7)
Let
1
1
j
n
j ij
i
j
m Y
n
=
=

and /
j j
P n M = ,
where M is the total number of training samples, m
j
denotes the mean projection
vector of class j and P
j
is a prior probability of class j.Then the between-class scat-
ter matrix S
b
, within-class matrix S
w
, and total population scatter matrix S
t
are
defned as
1
[ ( )][ ( )]
c
T
b j j j
j
S P m E Y m E Y
=
= - -
, (9.8)
1
[( )( ) ]
c
T
w j j j j
j
S P E Y m Y m
=
= - -
, (9.9)
{[ ( )][ ( )] }
T
t b w
S E Y E Y Y E Y S S = - - = + , (9.10)
where E( ) denotes the expectation value of vectors or matrices.
In order to construct the criterion function for class separability, we need to
transform the above matrices to numbers. The criteria should be larger when the
between-class scatter is larger or the within-class scatter is smaller. To this end, the
following function is constructed from Eqs. (9.8) and (9.10), which is a generaliza-
tion of classical linear discriminant analysis:
( )
( )
b
t
tr S
J
tr S
=
, (9.11)
where tr() denotes the trace of matrices.
According to Eq. (9.8), one obtains the following equation:
1
( ) [ ( )] [ ( )]
c
T
b j j j
j
tr S P m E Y m E Y
=
= - -
. (9.12)
is prohibited.
Substituting Eq. (9.7) into Eq. (9.12), one obtains
1
( ) [ ( ) ] [ ( ) ]
c
T
b j j j
j
tr S P A X E A X A X E A X
=
= - -
1
[ ( )] [ ( )]
c
T T
j j j
j
X P A E A A E A X
=
= - -
, (9.13)
where A
j
is the average image matrix of class j.
Defne the matrix below
1
1
[ ( )] [ ( )]
c
T
b j j j
j
S P A E A A E A
=
= - -
. (9.14)
The matrix S
b1
is called image between-class scatter matrix. It is obvious that S
b1

is an n n matrix. In a similar way, we can defne the following two matrixes:
1
1
[( ) ( ) ]
c
T
w j j j j
j
S P E A A A A
=
= - -
, (9.15)
1
{[ ( )] [ ( )]}
T
t
S E A E A A E A = - - . (9.16)
The matrix S
w1
is called image within-class scatter matrix and the matrix S
t1
is
called image total population scatter matrix. Accordingly, it is easy to verify that
S
w1
and S
t1
are n n matrices. In such a case, we transform Eq. (9.11) into the fol-
lowing form:
1
1
( )
T
b
T
t
X S X
J X
X S X
=
. (9.17)
In general, the problem of Eq. (9.17) can be solved by the following generalized
eigenvalue problem:
1 1 b t
S X S X = . (9.18)
It should be pointed out that the matrix S
t1
is nonsingular unless there is only
one sample for each class. As discussed in Yang, Yang, Frangi and Zhang (2003),
the eigenvector corresponding to maximum eigenvalue of Eq. (9.18) is taken as the
frst uncorrelated discriminant vector. According to Jings theory, the (d+1)th un-
correlated discriminant vector X
d+1
is the eigenvector corresponding to maximum
eigenvalue of the following eigenequation:
is prohibited.
1 1 b t
PS X S X =
, (9.19)
where
1
1 1
( )
T T
t t
P I S D DS D D
-
= - ,
1 2
( )
T
r
D X X X = , and I = diag(1,1,...,1).
In the following, an effective method for two-dimensional linear discriminant
analysis is described, which applies the Fisher criterion and statistical correlation
between extracted features. Assume that optimal 2D projection vectors X
1
, X
2
,...,X
d

are used for feature extraction. Let ( 1, , )
k k
Y AX i d = = . Thus, the image space
is transformed into the feature space. Let Y
i
and Y
j
be any two features. Then the
covariance between them is defned as
( ) ( ) ( ) ( ) ( )
1
cov ,
T
i i i i j j i t j
Y Y E Y E Y Y E Y X S X
(
= - - =

(9.20)
Accordingly, the correlation coeffcient between Y
i
and Y
j
is defned as
cov( , )
( , )
cov( , ) cov( , )
i j
i j
i i j j
Y Y
Y Y
Y Y Y Y
=

1
1 1
( ) ( )
T
i t j
T T
i t i j t j
X S X
X S X X S X
=
.

(9.21)
For the sake of discussion, let ( , ) ( , )
i j i j
Y Y f X X = . In a similar way, we select
the vector corresponding to the maximum value of Eq. (4.17) as the frst discrimi-
nant vector. Then the following optimization model is used to obtain the (d+1)th
discriminant vector, denoted as
max J(X),
2
1 1
min ( , ) f X X
2
min ( , )
d d
f X X
,
where
( , )
i i
f X X

1
1 1
T
i t
T T
i t i t
X S X
X S X X S X
= , i = 1,...,d. (9.22)
It is obvious that the correlation between X and X
i
(i = 1,...,d), namely Y and Y
i
(i =
1,...,d), is the lowest when
2
( , )
i r
f X X i = 1,...,d are zero. This new model shows
that the feature vector extracted by the (d+1)th discriminant vector has the lowest
correlation with those extracted by previous d discriminant vectors.
In order to deal with the above model, the model is further transformed into the
following equation:
is prohibited.
2
0
1
( ) ( ) ( , )
d
i i i
i
G X s J X s f X X
=
= -
, (9.23)
where 0( 0, , )
i
s i r = are weighting factors and
0
1
d
i
i
s
=
=
.
From Eq. (9.23), we can see that the smaller
2
( , )
i r
f X X is and the bigger J(X)
is, the bigger G(X) is. Therefore, it is necessary to obtain the (d+1)th discriminant
vector corresponding to the maximal value of G(X).
Substituting Eqs. (9.17) and (9.21) into Eq. (9.23), one can obtain
1
0
1
( )
T
b
T
t
X S X
G X s
X S X
=
2
1 1
1
1
( ) /( )
T T
d
i t i t i
i
T
i
t
X S X X S X
s
X S X
=
-
. (9.24)
From Eq. (9.24), it is straightforward to verify that for any nonzero constant ,
G(X) = G(X). In such a case, one adds the constraint to G(X) and the correspond-
ing model is denoted as
max G(X),
X
T
S
t1
X = 1. (9.25)
In order to further deal with Eq. (9.25), one can construct the following Lagrange
function in terms of the Lagrange multiplier , denoted by
1
( ) ( ) ( 1)
T
t
L X G X X S X = - - . (9.26)
Setting the partial derivative of L(X) with respect to X equal to zero, one ob-
tains
0 1 1 1 1 1
1
2 2 2 ( ) /( ) 0
d
T T
b t i t i i t i t i
i
s S X S X s S X X S X X S X
=
- - =
. (9.27)
Then one can obtain the following equation:
0 1 1 1 1
1
( ( ) /( ))
d
T T
b i t i i t i t i
i
s S s S X X S X S X
=
-

1 t
X S X = , (9.28)
From the above discussion, we obtain the following theorem.
Theorem 9.1 The (d + 1)th discriminant vector is the vector corresponding to
maximum eigenvalue of the following generalized eigenequation:
PX = S
t1
X, (9.29)
is prohibited.
where
0 1 1 1 1
1
( ) /( )
d
T T
b i t i i t i t i
i
P s S s S X X S X S X
=
= -
,
0
1
d
i
i
s
=
=
, s
i
0(i = 0,...,d).
Compared with Eq. (9.19), we can see that it is not necessary to use the matrix
inverse in Eq. (9.29). Moreover, we can directly apply previous results to compute
P in Eq. (9.28). Therefore, performing the 2DLDA method costs less computational
time than performing uncorrelated discriminant analysis. Applying Eq. (9.29), we can
obtain optimal discriminant vectors {X
1
,...,X
r
}. Then corresponding Fisher criterion
values can be obtained by Eq. (9.17). As is pointed out in Xu,Yang and Jin (2003),
the Fisher criterion value of Lius method (1993) corresponding to each discriminant
vectors is not smaller than that of the corresponding uncorrelated discriminant vec-
tor. One asks: does there exist a relationship of Fisher criterion values between the
2DLDA method and UIDA. To answer this question, we frstly give some related
knowledge on generalized eigenvalue problems. Assume that X
1
,...,X
n
are linear
independent eigenvectors of
1
1 1 t b
S S
-
corresponding to the eigenvalues
1 n
.
Let
1 1
{ , , }
r n
W span X X
+
= and
2 1
{ , , }
r
W span X X = . Then we obtain
1 1
max{ ( ) 0 }
r
J X X W
+
= , (9.30)
2
min{ ( ) 0 }
r
J X X W = . (9.31)
According to the above theories, we can obtain the following proposition.
Proposition 9.2 Let J
y
(X
i
)(i = 1,...,n) be Fisher criterion values obtained by uncor-
related image discriminant analysis and J
l
(X
i
)(i = 1,...,n) be Fisher criterion values
obtained by the 2DLDA method. Then we obtain
J
l
(X
i
) J
y
(X
i
) (i = 1,...,n). (9.32)
Proof. According to the theories of generalized eigenvalue problems, it is sure
that J
y
(X
i
) =
i
(i = 1,...,n). However, the discriminant vectors {X
1
,...,X
n
} using the
2DLDA method may not be a basis of R
n
. Assume that the rth discriminant vector
X
r1
is obtained using the 2DLDA method. If X
r1
W
2
, J
l
(X
r1
)

r
= J
y
(X
r
); if X
r1

W
1
, J
l
(X
r1
)

r + 1
= J
y
(X
r
)
r
=

J
y
(X
r
). Therefore, the proposition holds.
Proposition 9.2 states that Fisher criterion values of the 2DLDA algorithm must
not be smaller than those of corresponding uncorrelated discriminant vectors.
Since Fisher criterion and statistical correlation are applied to construct the
evaluation function, one guesses there exists the relationship between the 2DLDA
method and uncorrelated discriminant analysis. To this end, we further analyze the
proof of uncorrelated discriminant analysis which can be found in Jin, Yang, Tang
and Hu (2001). In deducing Eq. (4.19), the following equation is applied:
is prohibited.
1 1 1
1
2 2 0,
d
b t t i i
i
S X S X S X
=
- - =
(9.33)
where
1 1
2 /( )
T T
i i b i t i
X S X X S X =
.
Substituting
i
into Eq. (9.33), one can obtain
1 1 1 1
1
/( )
d
T T
b t i i b i t i
i
S X S X X S X X S X
=
-

1 t
S X =
. (9.34)
In what follows, the relationship between Eq. (9.34) and Eq. (9.29) is further
discussed.
Let
1 1 1 1
/( )
T T
i t i i b t i i t
c S X X S S X X S = ,
then one transforms Eq. (9.34) into the following form:
1 1 1 1
1
( /( ))
d
T T
b i t i i t i t i
i
S c S X X S X S X X
=
-

1 t
S X =
. (9.35)
Let
0
1
1
1
r
i
i
s
c
=
=
+
and
( )
1
1, ,
1
i
i r
i
i
c
s i r
c
=
= =
+
. .
In such a case, it is obvious that
0
1
r
j
j
s
=
=
, which satisfes the constraint in Eq.

(9.29). Therefore, Eq. (9.29) is a generalization of Eq. (9.19). In other words, if we
choose suitable parameters in Eq. (9.28), the solutions to Eq. (9.28) is equivalent
to the solutions to Eq. (9.19). As a result, one builds a bridge between the 2DLDA
algorithm and UIDA. That is, the relationship between Eq. (9.29) and Eq. (9.19) is
revealed. From the above discussion, we also fnd an effcient method for computing
the matrix P in Eq. (9.19), which doesnt need to use the matrix inverse.
9.2.3 t wo-dimensional l PP (2dl PP)
A. One-Sided 2DLPP
Since we focus on dealing with image matrices, it is useful to introduce image
projections before we begin. Let A
1
= A
M
be M image matrices, where A
i
is an m
n matrix. Then the image A is projected onto an n-dimensional vector u. That is,
the following transformation is adopted: Y
i
= A
i
u. It is obvious that the projection
vector Y
i
is an m-dimensional vector.
is prohibited.
As discussed in the standard spectral theory (Chung, 1997) for data represen-
tation vectors, a reasonable criterion for choosing a good map is to minimize the
following objective function
2
,
( )
i j ij
i j
y y W -
under approximate constraints. If

neighbor points are mapped far apart, the objective function suffers from a heavy
penalty by the weights. Applying the ideas, one can develop the following objective
function for image matrices, denoted by
2
1
2
,
( ) / 2
i j ij
i j
f u Au A u W = -
. (9.36)
Applying some algebra operations, we deal with f
1
(u).
2
1
2
,
( ) / 2
i j ij
i j
f u Au A u W = -

,
1
( ) ( )
2
T
i j i j ij
i j
Au A u Au A u W = - -
,
1
( ) ( )
2
T T
i j ij i j
i j
u A A W A A u = - -
.
For the sake of simplicity, let us defne the following matrix
G
w
=
,
1
( ) ( )
2
T
i j ij i j
i j
A A W A A - -
.
It is obvious that G
w
is an n n matrix. It is easy to obtain the matrix G
w
if W
is known.
G
w
is further presented by
G
w
=
,
1
( )
2
T
i ij i
i j
A W A

,
1
( )
2
T
j ij j
i j
A W A +

,
1
( )
2
T
i ij j
i j
A W A -

1 1
( )
T
m
F D I F =
1 1
( )
T
m
F W I F -
where
1 1
[ ]
T T
s
F A A = , D is a diagonal matrix whose diagonal elements is the column
sum of W, that is,
ii ij
j
D W =
, I
m
is an m m identity matrix, denotes Kronecker
product. In such a case, let us defne the following matrix: L = D I
m
W I
m
,
then L is a generalized Laplacian matrix.
As discussed in He and Niyogi (2003), we add the constraint on the objective
function f
1
(u). The corresponding constraint is denoted as follows:
1 1
( ) 1
T T
m
u F D I F u = .
is prohibited.
In all, the optimization problem can be transformed into
1 1 1
min ( )
T T
f u u F LF u = , (9.37)
subject to
1 1
( ) 1
T T
m
u F D I F u = .
The vector u which minimizes Eq. (9.37) can be obtained by solving the following
generalized eigenvalue problems
1 1
T
F LF u =
1 1
( )
T
m
F D I F u . It is straightforward
to verify that the optimal u is the eigenvector corresponding to the smallest eigen-
value. In some cases, it is not enough to obtain an optimal vector u. To this end, the
eigenvectors which correspond to the frst l smallest eigenvalue can be obtained.
B. The Relationship between 2DPCA and 2DLPP
Firstly, let us recall the 2DPCA. In general, the idea of 2DPCA is to solve the fol-
lowing eigenvalue problem:
G
t
u = u
where
1
( ) ( ) /
M
T
t i i
i
G A A A A M
=
= - -
and
1
/
M
i
i
A A M
=
=
.
Substituting A into G
t
and applying some algebra operations, we have
1 1
1 1
( )
T
t m M M
G F I I P I F
M M
= - ,
where P is an m m matrix whose entries are 1.
The 2DPCA algorithm is to obtain the frst l eigenvectors which correspond to the
frst l largest eigenvalues. This method can preserve the global structure. However,
the 2DLPP algorithm is to obtain the frst l eigenvectors which correspond to the
frst l smallest eigenvalues. It can preserve local information. It is interesting to note
1 1
( )
T
m
F D I F
corresponds to weighed 2DPCA when 0 A = . This further verifes
that effectiveness of Eq. (9.37), namely, Eq. (9.37) can preserve the local neighbor
structure and simultaneously maximize the global structure.
C. A Generation of 2DLPP
In this section, we further extend 2DLPP. More specifcally, one can apply two-
sided projections on image matrices. By virtue of some ideas in Ye, Janardan and
Li (2004), one considers the following l
1
l
2
-dimensional space L R, where L is
is prohibited.
spanned by
1
1
{ , , }
l
u u and R is spanned by
1
1
{ , , }
l
v v . Then the following matrices
are defned:
1
1
[ ]
l
L u u = and
2
1
[ ]
l
R v v = . It is easy to show that L should be an m
l
1
matrix and R should be an n l
2
matrix. Accordingly, the image matrix A is
projected on the space L R by L
T
AR in such a case.
Similar to the discussion in Section 9.2.3, the following objective function is
adopted:
2
2
,
1
( , )
2
T T
i j ij
F
i j
f L R L AR L A R W = -
.
Applying some algebra operations, one further deals with f
2
(L,R).
2
2
,
1
( , ) (
2
T T
i j ij
F
i j
f L R L AR L A R W = -
,
1
( ) ( )
2
T T T
i j ij i j
i j
L AR A R W R A A L = - -
2
2 1 1 2
( )
T T T
l
L F H D I H F L
2
2 1 1 2
( )
T T T
l
L F H W I H F L -
1
T R
L S L = ,
where F
2
= [A
1
...A
s
], H
1
= diag(R,...,R),
1
R
S

2
2 1
(
l
F H D I =

2
1 2
)
T T
l
W I H F -
.
Similarly, one adds the following constraint:
2
2 1 1 2
( )
T T T
t l
D L F H D I H F L =
1
2
T R
l
L S L I = = .
In such a case, the optimization problem can be reduced to
2
,
argmin ( , )
L R
f L R (9.38)
subject to
1
2
T R
l
L S L I = .
It is obvious that this optimization problem involves two transformation ma-
trices, L and R. It seems that it is very diffcult to solve this optimization problem.
Fortunately, we note that the optimization problem will minimize tracef
2
(L,R) and
maximize trace(D
t
) for a fxed R. More specifcally, under the assumption that R
is the constant matrix, the optimal L can be obtained by solving the generalized
eigenvalue problem
1 2
R R
S u S u = . The optimal L consists of the frst l

eigenvectors
which correspond to the frst l
1
smallest eigenvalues of
1
2 1
( )
R R
S S
-
. For a fxed L , it
is necessary to minimize tracef
2
(L,R) and simultaneously maximize trace(D
t
) in a
similar manner. Note that trace(AB) = trace(BA), for any two matrices A and B. In
is prohibited.
such a case, tracef
2
(L,R) and trace(D
t
) can be described as follows:
tracef
2
(L,R)
,
1
( ) ( )
2
T T T
i j ij i j
i j
R A A LW L A A R = - -
=
1 1
1 2 2 1
( ( ) )
T F T
l l
trace R F H D I W I H F R -
,
where H
2
= diag(L,...,L),
1
L
S
1
1 2
(
l
F H D I =
1
2 1
)
T T
l
W I H F - .
trace(D
t
)
1
1 2 2 1
( ( ) )
T T T
l
trace R F H D I H F R =
2
( )
T L
trace R S R = .
Likewise, the optimal R can be obtained by solving the following eigenvaule
problem
1 2
L L
S v S v = . Accordingly, R consists of the frst l
2
eigenvectors of
1
2 1
( )
L L
S S
-

which corresponds to the frst l
2
smallest eigenvalues.
From the above discussion, we also obtain an iterative algorithm for solving
Eq. (9.38).
The iterative algorithm is briefy described as follows.
The iterative algorithm of 2DLPP
Step1: Choose R
0
and compute
2
R
S and
1
R
S .
Step2: Compute the frst l
1
eigenvectors of
1
2 1
( )
R R
S S
-
and obtain L.
Step3: Substitute L into
2
L
S and
1
L
S .
2
eigenvectors of
1
2 1
( )
L L
S S
-
and obtain R.
Step5: Let R
0
= R, goto step 1.
It should be pointed out that the above algorithm adopts similar strategies with
Yes 2DLDA. A more detailed discussion on 2DLDA can be found in Ye, Janardan
and Li (2004). The difference between 2DLDA and the iterative algorithm is that
the latter can be performed in supervised or unsupervised manner while 2DLDA
is supervised. Here a natural way to initialize R
0
is to use R
0
= (I
l2
, 0)
T
which is
discussed in Ye, Janardan and Li (2004). It is clear that the iterative algorithm is
required to run several times for obtaining L and R, which results in the computa-
tional diffculty. However, it is of interest to note that the following lemma can be
applied to simplify the above iterative algorithm.
Lemma9.3(Schott,1997).Let A be m m nonnegative defnite matrix and
B be an m m positive defnite matrix. If G is any m h matrix with full column
rank, then
1 1
(( ) ) ( )
T T
h
trace G BG G AG trace B A
- -
,
where trace(B
1
A)
h
denotes the sum of the frst h largest eigenvalues.
is prohibited.
One can note that the minimization problem in Eq. (9.38) can be transformed
into the maximization problem:
1
1 2
max ( ) ).
T R T R
trace L S L L S L
-
Applying Lemma 9.3 , one has
1
1 2
( ) )
T R T R
trace L S L L S L
-

1
1 2
(( ) ).
R R
trace S S
-
In order to solve
1
1 2
max (( ) )
R R
trace S S
-
, it is necessary to maximize
2
( )
R
trace S
and minimize
1
( )
R
trace S . In such a case,
2
( )
R
trace S and
1
( )
R
trace S can be further
written
2 1 1
( ) ( ( ) )
R T T
n
trace S trace R F D I F R =
= trace(R
T
S
2
R),
1
( )
R
trace S
1 1
( ( ) )
T T
n n
trace R F D I W I F R = -
= trace(R
T
S
1
R).
In a similar way, we obtain R by solving the generalized eigenvalue problem S
1
v
= S
2
v. R consists of the frst l
2
eigenvectors of (S
2
)
1
S
1
which corresponds to the frst
l
2
smallest eigenvalues. From the above discussion, we can propose the following
algorithm for obtaining L and R.
Non-Iterative Algorithm of 2DLPP
Step1: Obtain S
1
and S
2
2
eigenvector (S
2
)
1
S
1
and obtain R.
Step3: Obtain
2
R
S and
1
R
S .
2
eigenvectors of
1
2 1
( )
R R
S S
-
and obtain L.
It is clear that the above algorithm is an non-iterative method for obtaining L
and R. Compared with the iterative algorithm, non-iterative algorithm has higher
computational effciency in general. As a matter of fact, non-iterative algorithm is
two-level 2DLPP from our observations.
9.2.4 Multilinear LDA
Assume that
1 N
I I
i
A

, for i = 1,...,M, are M tensors in the dataset. These M ten-
sors (samples) consist of c classes, with each class having n
i
samples.
Let
1
1
i
n
i j
j
i
M A
n
=
=

is prohibited.
denote the mean tensor of the ith class, 1 i c and
1
1 M
j
j
M A
M
=
=

be the total mean tensor of all t raining samples. Given
1 1
(1) I R
U

,
2 2
(2)
,
I R
U

,
( )
N N
I R N
U

, we consi der t he f ol l owi ng s pa ce
(1) ( ) N
U U ., the tensor A
i
is projected onto
(1) ( ) N
U U , denoted by
i
A

(1) ( )
1
(( ) ( )
T N T
i N
A U U =
Similar to classical LDA, multilinear LDA is to fnd the optimal projections
1 1
(1) I R
U

,
2 2
(2)
,
I R
U

,
( )
N N
I R N
U

such that the between-class scatter is
maximized and the within-class scatter is minimized simultaneously. To this end,
we defne the following between-class distance and within-class distance in terms
of the Frobenius-norm of tensors:
2
1 1
i
n
c
w j i
F
i j
D A M
= =
= -
, (9.39)
2
1
c
b i i
F
i
D n M M
=
= -
. (9.40)
Similarly, one can defne the between-class distance and within-class distance
in the projection space U
(1)
... U
(N)
as follows:
2
(1) ( )
1
1 1
( ) ( ) ( )
i
n
c
T N T
w j i N
F
i j
D A M U U
= =
= -
, (9.41)
2
(1) ( )
1
1
( ) ( ) ( )
c
T N T
b i i N
F
i
D n M M U U
=
= -

. (9.42)
It seems that it is very diffcult for us to deal with Eqs. (9.41) and (9.42). However,
the matrix representation of tensors provides an effective strategy for simplifying
Eqs. (9.41) and (9.42). Note that two tensors satisfying the following relationship
(1) (2) ( )
1 2
( ) ( ) ( )
T T N T
N
S A U U U = (9.43)
can be represented in the following matrix form:
( ) ( 1) ( ) (1) ( 1)
( ) ( )
( ) ( )
n T n N n
n n
S U A U U U U
+ -
= . (9.44)
is prohibited.
It is not diffcult to verify that
2
2
( ) n
F
F
A A = , for n = 1,...,N. In such a case, we
substitute Eq. (9.44) into Eqs. (9.41) and (9.42) and obtain
2
( ) ( 1) ( )
( )
1 1
( ) ( ) ( )
i
n
c
n T n N
w j i n
F
i j
D U A M U U
+
= =
= -
, (9.45)
2
( ) ( 1) ( )
( )
1
( ) ( ) ( )
c
n T n N
b i i n
F
i
D n U M M U U
+
=
= -
. (9.46)
In general, in order to obtain the optimal projections U
(1)
, U
(2)
,...,U
(N)
, it is
necessary to maximize D
b
and to minimize D
w
simultaneously. We note that it is
very diffcult to fnd the optimal projections simultaneously. To this end, we de-
velop an iterative algorithm for obtaining the optimal projections U
(1)
, U
(2)
,...,U
(N)

like the algorithm of HOSVD or 2DLDA. Furthermore, we can see that for fxed
(1) ( 1) ( 1) ( )
, , , , ,
n n N
U U U U
- +
Eqs. (9.41) and (9.42) can be further represented as
follows:
2
( ) ( 1) ( )
( )
1 1
( ) ( ) ( )
i
n
c
n T n N
w j i n
F
i j
D U A M U U
+
= =
= -

( ) ( ) ( )
( )
n T n n
b
trace U S U
where
( ) ( 1) ( ) ( 1) ( )
( ) ( )
1 1
( ) ( )[( ) ( )]
i
n
c
n n N n N T
w j i n j i n
i j
S A M U U A M U U
+ +
= =
= - -
,
and
2
( ) ( 1) ( )
(1)
1
( ) ( ) ( )
c
n T n N
b i i
F
i
D n U M M U U
+
=
= -

( ) ( ) ( )
(( ) )
n T n n
b
trace U S U =
where
( ) ( 1) ( ) ( 1) ( )
( ) ( )
1
[( ) ( )][( ) ( )]
c
n n N n N T
w i i n i n
i
S n M M U U M M U U
+ +
=
= - -
.
Similar to the 2DLDA algorithm, it is not diffcult to obtain the optimal projection
U
(n)
. More specifcally, for fxed U
(1),...,
U
(n-1)
U
(n+1)
U
(N)
,

the optimal projection U
(n)
can
be obtained by solving the generalized eigenvalue problem
( ) ( ) ( ) ( ) n n n n
b w
S U S U = . In
other words, U
(n)
consists of the frst R
n
eigenvectors which correspond to the frst
R
n
largest eigenvalues of
( ) 1 ( )
( )
n n
w b
S S
-
. In short, as a summary of the above discus-
sion, the algorithm is shown Figure 9.1.
From the algorithm, one can note the following facts: when A
i
(i = 1,...,M) are
vectors, the algorithm is degenerated into classical LDA; when A
i
(i = 1,...,M) are
matrices, the algorithm is degenerated into 2DLDA. Therefore, in some sense, the
2DLDA method unifes current LDA and 2DLDA. It is obvious that the 2DLDA
is prohibited.
method can directly deal with multidimensional data. Since the algorithm is iterative,
theoretically, the solution is local optimal. Like the HOSVD algorithm, probably it
has several optimal solutions. In addition, we can see the algorithm involves setting
initial values and judging the convergence. Therefore, it is necessary to discuss these
two problems. By virtue of some ideas of HOSVD, we give the following strategy
to obtain the initial values.
From Eqs. (9.39) and (9.40), we can obtain
2
1 1
i
n
c
w j i
F
i j
D A M
= =
= -

( ) ( )
1 1
[( ) ][( ) ]
i
n
c
T
j i n j i n
i j
trace A M A M
= =
- -

( )
( )
w n
trace S
where
Figure 9.1. The iterative algorithm for HDLDA
In:
1 N
I I
i
A

, for i = 1,...,M
Out: U
(n)
, n =1,...,N;
1 N
R R
i
A

, for i =1,...,M.
1. initial values:
( )
0
n
U
(2 n N));
2. compute the mean tensor M
i
of class i and the globle mean tensor M;
3. iterative until convergence
(i) compute E
1
=
(2) ( )
( )
N
k k
U U ,
1 (1) 1 (1) 1
1
( ) [( ) ][( ) ]
c
T
b i i i
i
S n M M E M M E
=
= - -
(ii) obtain
( )
1
N
k
U
+ consisting of the frst R
1
eigenvectors of
1
1 1
[( ) ] ( )
w b
S S
-
.
(i) compute E
N
=
(1) ( 1)
( )
N
k k
U U
-
,
( ) ( )
1 1
( ) [( ) ][( ) ]
i
c n
T
w N j i N N j i N N
i j
S A M E A M E
= =
= - -

( ) ( )
1
( ) [( ) ][( ) ]
c
T
b N i i N N i N N
i
S n M M E M M E
=
= - -
(ii) obtain
( )
1
N
k
U
+ consisting of the frst R
N
eigenvectors of
1
[( ) ] ( )
w N b N
S S
-
.
4. Obtain converged values U
(1)
,...,U
(N)
.
5. Compute
(1) ( )
1
( ) ( ) ( )
T N T
i i N
A A U U = , for i = 1,...,M.
is prohibited.
( ) ( ) ( )
1 1
( ) [( ) ][( ) ]
i
n
c
T
w n j i n j i n
i j
S A M A M
= =
= - -
. (9.47)
2
1
c
b i i
F
i
D n M M
=
= - =

( ) ( )
1
[ ( ) ][( ) ]
c
T
i i n i n
i
trace n M M M M
=
- -

( )
( )
b n
trace S = ,
where
( ) ( ) ( )
1
( ) [( ) ][( ) ]
c
T
b n i i n i n
i
S n M M M M
=
= - -
. (9.48)
In such a case, we obtain the initial values
( )
0
n
U consisting of the frst R
n
eigen-
vectors of
1
( ) ( )
[( ) ] ( )
w n b n
S S
-
. As discussed in He and Niyogi (2003), the main reason
for choosing these initial values is that the optimal solution could be obtained by
few iterations. In a similar way, the stop criterion used in the algorithm can take the
form, for example,
2
( ) ( )
1
n n
k k n
F
U U R
+
- (1 n N), where is a positive number.
However, it is of interest to note that Lemma 9.3 can be applied to obtain a non
iterative algorithm for HDLDA.
From Lemma 9.3, we have
( ) ( ) ( ) 1 ( ) ( ) ( )
( ) ( )
n n n n T n n
w b
trace U S U U S U
-

( ) 1 ( )
( )
n n
w b
trace S S
-
. (9.49)
From Eq. (9.49), we can see that it is necessary to minimize
( )
( )
n
w
trace S and
maximize
( )
( )
n
b
trace S simultaneously. Note that
( )
( )
n
w
trace S

2
(1) ( 1) ( 1) ( )
1 1 1
1 1
( ) ( ) ( ) ( ) ( )
i
n
c
T n T n T N T
j i n n N
F
i j
A M U U U U
- +
- +
= =
-

(9.50)
( )
( )
n
b
trace S

2
(1) ( 1) ( 1) ( )
1 1 1
1
( ) ( ) ( ) ( ) ( )
c
T n T n T N T
i i n n N
F
i
n M M U U U U
- +
- +
=
= -

(9.51)
Applying the matrix form of tensors from Eq. (9.50) and (9.51) and applying
lemma 9.3, one can further reduce one projected matrix in Eq. (9.50) and (9.51). In
a similar way, we can reduce the projected matrices until only one projected matrix
is preserved. Without loss of generality, we assume that U
(1)
is preserved. In other
words, there exist the following equations:
2
( ) (1)
/2, , 1
1 1
( ) ( ) ( )
i
n
c
n T
w N j i
F
i j
trace S A M U
= =
= -
.
(9.52)
is prohibited.
2
( ) (1)
/2, , 1
1
( ) ( ) ( )
c
n T
b N i i
F
i
trace S n M M U
=
= -
.
. (9.53)
From Eq. (9.52) and (9.53), it is necessary to minimize
( )
/2, ,
( )
n
w N
trace S
.
and
maximize
( )
/2, ,
( )
n
b N
trace S
.
simultaneously. In such a case, it is not diffcult to
verify that the optimal projected matrix U
(1)
consists of the frst R
1
eigenvectors
of
1
(1) (1)
[( ) ] ( )
w b
S S
-
, where
(1)
( )
w
S and
(1)
( )
b
S are defned in Eq. (9.47) and Eq. (9.48).
Furthermore, we apply lemma 9.3 to obtain other projected matrices until the last
projected matrix is achieved. The corresponding algorithm is described in Figure
9.2.
From Figure 9.2, it is obvious the algorithm is non-iterative. Compared with
the algorithm in Figure 9.1, the non-iterative algorithm has higher computational
effciency in general.
Figure 9.2. The non-iterative algorithm for HDLDA
In:
1 N
I I
i
A

, for i = l,...,M
Out: U
(n)
n = 1,...,N;
1 N
R R
i
A

, for i = 1,...,M.
1 compute the mean tensor M
i
of class i and the global mean tensor M;
2 (i) compute
1 (1) (1)
1 1
( ) [( ) ][( ) ]
i
c n
T
w j i j i
i j
S A M A M
= =
= - -

(ii) obtain U
(1)
consisting of the frst R
1
eigenvectors of
1
1 1
[( ) ] ( )
w b
S S
-
.
(i) compute E
1
= U
(1)
I
(3)
I
(N)

2 (2) 1 (2) 1
1 1
( ) [( ) ][( ) ]
i
c n
T
w j i j i
i j
S A M E A M E
= =
= - -

2 (2) 1 (2) 1
1
( ) [( ) ][( ) ]
c
T
b i i i
i
S n M M E M M E
=
= - -
(ii) obtain U
(2)
2
eigenvectors of
1
2 2
[( ) ] ( )
w b
S S
-
.
(i) compute E
N-1
= U
(1)
... U
(N-1)
,

( ) ( )
1 1
( ) [( ) ][( ) ]
i
c n
T
w N j i N N j i N N
i j
S A M E A M E
= =
= - -

( ) ( )
1
( ) [( ) ][( ) ]
c
T
b N i i N N i N N
i
S n M M E M M E
=
= - -
(ii) obtain U
(N)
N
eigenvectors of
1
[( ) ] ( )
w N b N
S S
-
.
3 Compute
(1) ( )
1
( ) ( ) ( )
T N T
i i N
A A U U =
, for i = 1,...,M.
is prohibited.
9.3.1 t he Performance of 2dld A on f ace Images
In order to verify and test the effectiveness of 2DLDA (Section 9.2.2), experiments
are made on the ORL face image database (please refer to Section 3.2.3). Each face
image is downsampled to 28 23 pixels to reduce the computational complexity.
In all experiments, we apply the frst fve images of each subject for training and
others are for testing. Namely, 200 samples are used for training and 200 samples
are used for testing.
The frst set of experiments is used for showing the effectiveness of the number
of features. The parameters in 2DLDA (Section 9.2.2) are set as follows: s
i
= 1/(r
+ 1) (i = 0,...,r). At the same time, 2DLDA is compared with uncorrelated image
discriminant analysis (UIDA) and Lius method (LM) (1993). In addition, the near-
est neighbor classifer is adopted for classifcation due to the simplicity. Table 9.1
shows the classifcation performance of several methods when the number of feature
vectors varies from 1 to 6. From Table 9.1, we can surprisedly see that the results
of 2DLDA are the same as those of UIDA. One can also note that the classifcation
performance of the 2DLDA algorithm is superior to that of Lius method.
In the second set of experiments, similarly, the nearest neighbor classifer is used
for classifcation and the parameters are set as follows: s
i
= 1/(r + 1) (i = 0,...,r). In
such a case, the execution time for feature extraction and classifcation of several
methods are compared, which is shown in Table 9.2. As can be seen from Table 9.2,
the methods based on image matrices including UIDA, LM and 2DLDA need less
time than Eigenfaces or Fisherface. There is no remarkable difference in time for
UIDA, LM and the 2DLDA method. Since both Eigenfaces and Fisherfaces need to
convert image matrices into vectors in the process of recognition, the classifcation
time of these two methods are more than that of 2D linear discriminant analysis.
Moreover, 2D linear discriminant analysis is superior to classical linear discriminant
analysis in terms of the computational effciency for feature extraction.
In the third set of experiments, similar experimental conditions are set. Based
on this, the classifcation performance of 2DLDA is compared with other methods
Table 9.1. Recognition accuracy of several methods on the ORL face database
The number of features 1 2 3 4 5 6
UIDA 0.8350 0.8900 0.8850 0.8850 0.8900 0.8900
LM 0.8350 0.8450 0.8650 0.8550 0.8550 0.8500
Ours 0.8350 0.8900 0.8850 0.8850 0.8900 0.8900
is prohibited.
including Fisherfaces , Eigenfaces, ICA , uncorrelated image discriminant analysis
(UIDA), Lius method (1993) and direct recognition method (DRM). The detailed
experimental results are listed in Table 9.3. From Table 9.3, we can see that 2DLDA
is better than other methods except for uncorrelated image discriminant analysis
in recognition rates. To our surprise, the results of 2DLDA are the same as those
of uncorrelated image discriminant analysis.
In the fourth set of experiments, we discuss the Fisher criterion values of several
methods such as Lius method (1993) and uncorrelated image discriminant analy-
sis. Firstly, we set the parameters as follows: s
i
= 1/(r + 1) (i = 0,...,r). In addition,
we also discuss another case, namely s
0
= 0.0001 and s
i
= 1/r (i = 1,...,r). In such a
case, we think that the Fisher criterion plays an insignifcant role and the statistical
correlation between feature vectors plays an important role in feature extraction,
denoted by (2). From Table 9.4, it is obvious that the Fisher criterion value of UIDA
corresponding to each feature vector is the smallest in all methods. From Table
9.3, we know the classifcation performance of the 2DLDA algorithm is superior
to that of Lius method. This means the Fisher criterion value is not an absolute
criterion for measuring the discriminatory power of discriminant vectors. We
also fnd that the classifcation performance is not superior to that of uncorrelated
image discriminant analysis in the second case, which also shows that statistical
correlation is not an absolute criterion for measuring the discriminatory power of
discriminant vectors. Therefore, in order to obtain powerful discriminant vectors,
it is necessary to combine Fisher criterion values and statistical correlations among
feature vectors.
Table 9.2. the execution time for feature extraction and classifcation of fve meth-
ods
Methods Eigenfaces Fisherfaces Ours LM UIDA
Dimension 1200 139 322 323 322
The time for extraction (s) 27.85 97.14 4.03 7.18 4.23
Classifcation time (s) 59.51 34.46 3.96 5.81 4.01
The total time (s) 87.36 131.60 7.99 12.99 8.24
Table 9.3. Recognition rates (%) of several methods
Methods Eigenfaces Fisherfaces LM UIDA ICA DM Ours
Dimension 1200 139 323 322 140 3228 322
Recognition rates 88.0 86.0 86.5 89.0 84.2 83.5 89.0
is prohibited.
9.3.2 t he Performance of 2dl PP on f aces Images
In order to verify and test 2DLPP described in Section 9.2.3. The experiments are
made on several data sets. From iterative and non-iterative algorithms, we can see
that two important parameters l
1
and l
2
are involved. For the sake of simplicity, l
1

and l
2
are set to the common value l in the following experiments. The samples
belonging to the same class are considered to be closed. The weights n our experi-
ments are set
/
i j
x x t
ij
W e
- -
= . Since we focus on feature extraction by using 2DLPP,
the nearest neighbor classifer is adopted.
Face images are obtained from Olivetti Research Lab. Each face image is downs-
ampled to 28 23 pixels to reduce the computational complexity. Let us check the
effect of the number of iterations. To this end, an experiment is carried out by using
randomly chosen 5 images per class for training, the rest of image for testing. The
parameter l is set to 10. Figure 9.3 denotes experimental results, where the x-axis
denotes the number of iterations, and the y-axis denotes the recognition perfor-
mance. As can be seen from Figure 9.3 (a), it is necessary to use several iterations
to obtain good experiments results. In the meantime, we compare the performance
of non-iterative algorithm (A1) with that of the iterative algorithm (A2) when the
number of features changes in the interval of [1,16], where A2 is applied 10 itera-
tions. Figure9.3 (b) denotes the experiments results. From Figure 9.3 (b), we can see
that the recognition performance increases with the increase of features in general
case. Non-iterative algorithm and the iterative algorithm obtain almost similar
results. The smaller difference between may result from the precision of calcula-
tion. In addition, we compare our basic 2DLPP with LPP in terms of the running
time for feature extraction. The running time of 2DLPP for feature extraction is
11.45s, while LPP takes 16.45s to extract features. It is obvious that 2DLPP is faster
than LPP in terms of computational effciency for feature extraction. Finally, we
compare our algorithms with LPP with the change of features. In this experiment,
the leave-one-out strategy is adopted. Figure 9.3 (c) lists the experimental results
with the increase of l. As observed in Figure 9.3 (c), the recognition performance of
2DLPP is superior to that of classical LPP when the number of features is small. But
Table 9.4. Fisher criterion values of several methods
J(X
i
) X
1
X
2
X
3
X
4
X
5
X
6

UIDA 0.8667 0.8315 0.6562 0.5804 0.5012 0.4725
LM 0.8667 0.8399 0.7479 0.6757 0.6498 0.5746
Ours (1) 0.8667 0.8315 0.6562 0.5804 0.5012 0.4725
Ours (2) 0.8667 0.8315 0.6563 0.5806 0.5014 0.4722
is prohibited.
Figure 9.3. Experimental results on the ORL database (a) the effect of iterations
for A1 (b) the recognition performance of A1 and A2 with the change of features.(c)
the recognition performance of several methods with the change of feature by the
leave-one-out strategy.
0 2 4 6 8 10 12 14 16 18 20
0. 86
0. 865
0. 87
0. 875
0. 88
0. 885
0. 89
0. 895
0. 9
Number of iterations
C
o
r
r
e
c
t

r
a
t
e
s

0 2 4 6 8 10 12 14 16 18 20
0. 1
0. 2
0. 3
0. 4
0. 5
0. 6
0. 7
0. 8
0. 9
Number of features
C
o
r
r
e
c
t

r
a
t
e
s
A1
A2
(a)
(b)
continued on following page
is prohibited.
extended 2DLPP is inferior to classical LPP with respect to accuracy. The highest
recognition rate of 2DLPP is 97.50%, while LPP reaches 96.75%.
9.3.3 t he Performance of 2DLPP on handwritten Character
database
The used dataset of handwritten numeral characters are from the UCI data set re-
pository. For the description of the dataset, please refer to Section 8.3.2. Using the
same procedure as shown in Section 8.3.2, we produce 256-dimensional features
for each character image. We also choose only 100 samples from each class to
constitute training samples and others are used for testing. As discussed in ORL
experiments, the frst experiment is made to show the effectiveness of iterations.
In this experiment, the parameter l is set to 8. Figure 9.4 (a) shows the recognition
performance with the change of iterations. From Figure 9.4 (b), it is obvious that
it is not necessary to take several iterations to obtain good experimental results.
However, accuracy curve is unstable with the change of iterations. In a similar
way, we compare our algorithms including extended 2DLPP (non-iterative (A2)
and iterative (A1)) with classical LPP in terms of the change of features. Figure
9.4 (b) denotes the recognition performance with the increase of features. As ob-
served in Figure 9.4 (b), the performance of all the algorithms increases with the
increase of features in general cases. It is noted that 2DLPP obtains almost similar
experimental results with LPP, but is superior to extended 2DLPP in some cases.
0 5 10 15 20 25 30 35 40 45 50
0. 1
0. 2
0. 3
0. 4
0. 5
0. 6
0. 7
0. 8
0. 9
1
Number of features
C
o
r
r
e
c
t

r
a
t
e
s
LP P
2DLP P
A2
(c)
Figure 9.3. continued
is prohibited.
Figure 9.4. Experimental results on handwritten numeral characters (a) effect of
A1 in terms of the number of iterations (b) the effect of several methods in terms
of the change of features.

0 2 4 6 8 10 12 14 16 18 20
0. 932
0. 934
0. 936
0. 938
0. 94
0. 942
0. 944
0. 946
0. 948
Number of iterations
A
c
c
u
r
a
c
y

0 2 4 6 8 10 12 14 16 18 20
0. 1
0. 2
0. 3
0. 4
0. 5
0. 6
0. 7
0. 8
0. 9
1
Number of features
a
c
c
u
r
a
c
y
A2
2DLP P
A1
LP P
(a)
(b)
Moreover, the results of A2 are almost the same as those of A1. A2 is more effcient
than A1 due to its non-iterations. We also note that Extended 2DLPP needs less
coeffcients that 2DLPP.
is prohibited.
9.3.4 t he Performance of multilinear ld A
In order to test multilinear LDA, the experiments are made on color face images
which can be obtained by the University of Essex (UK) face database http://cswww.
essex.ac.uk/mv/ allfaces/index.html. The images in the Essex database are JPEG-
compared 24-bit color images. The size of each color face image is 200 180. The
subjects represent a wide distribution of ethnic group and both genders, but nearly
all subjects are between 18 and 20 years old. It is obvious that each color face im-
age corresponds to a three-order tensor. Hence, it is easy to apply the multilinear
method to this database. In order to reduce the computational cost, we only use
a subset of this face dataset. We select 40 persons with each person having 10
color images. In other words, the total number of samples is 400. Meanwhile, we
subsample the original image to 50 45-size images. In such a case, we randomly
choose 5 samples from each class to form the training samples and other samples
are used for testing. In order to reduce variation, the experimental results reported
in this paper are averaged over ten runs. For the sake of simplicity, we adopt the
Frobenius-norm as distance measure and consider the sample with minimal distance
as recognition samples. The frst set of experiments is used for showing the effect
of the number of iterations. Moreover, we compare the multilinear LDA method
with HOSVD. Figure 9.5 shows the experimental results, where the x-axis denotes
Figure 9.5. Recognition accuracy with the change of the number of iterations
is prohibited.
the number of iterations, and the y-axis denotes the recognition performance. In
this experiment, we set R
1
= R
2
= 10 and R
3
= 2. From Figure9.5, we can see that
the recognition performance is not changed violently when the number of iterations
is bigger than some value. It is clear that the multilinear LDA method can obtain
completive results with HOSVD. We also note that it is unnecessary for too many
iterations to obtain good recognition results.
The next experiment is used to show the effect of reduced dimensions. Since the
third dimension of tensors is small, we set R
3
= 2. Meanwhile, similar to 2DLDA,
we set R
1
= R
2
= d. Figure 9.6 shows the recognition performance with the change
of d. From Figure9.6, we note that the recognition performance frstly increases and
then decreases. It is clear that the multilinear LDA method obtains slightly better
results than HOSVD in terms of classifcation performance. Moreover, we note that
the non-iterative HDLDA (NHDLDA) method obtains almost the same results with
HDLDA. It shows that it may adopt the non-iterative method for solving HDLDA
in order to obtain the computational effciency in real applications.
9.4 summAr y And dI scuss Ion
In recent years, tensor discriminant analysis has attracted considerable interest from
researchers. In this chapter, we mainly describe several variants of classical LDA
including 2DLDA, 2DLPP and HOLDA after classical LDA is given.
Figure 9.6. Recognition accuracy with the change of the number features
is prohibited.
First, we describe two-dimensional linear discriminant analysis for image feature
extraction. It directly utilizes image matrices to construct the Fisher criterion func-
tion, which does not need to convert image matrices into high-dimensional vectors
such as classical LDA. Then discriminant vectors are obtained by maximizing Fisher
criterion functions and minimizing statistical correlation between extracted features.
Since the size of image matrices is much smaller that of vectors, the execution time
of two-dimensional linear discriminant analysis for feature extraction is much less
than that of traditional linear discriminant analysis. Moreover, we demonstrate that
the Fisher criterion values of two-dimensional linear discriminant analysis are no
less than the Fisher criterion values of uncorrelated discriminant vectors. In addi-
tion, the feature vectors obtained by two-dimensional linear discriminant analysis
are the same as those obtained by uncorrelated discriminant vectors in some con-
dition. It should be pointed out that the 2DLDA method requires more coeffcients
for feature extraction than LDA. In other words, the 2DLDA method needs more
storage space than the classical LDA method, which is one of disadvantages of 2D
linear discriminant analysis.
Second, two dimensional locality preserving projections are described. These
algorithms are based on image matrices instead of vectors. As a result, 2DLPP is
required to solve eigen-decomposition with small sizes, while LPP needs to solve
eigen-decomposition with large sizes. Moreover, 2DLPP can be performed in un-
supervised or supervised or semi-supervised manner. Similar to LPP, 2DLPP can
preserve the local neighborhood structure as possible. However, it is worthwhile to
point out that 2DLPP needs much more storage requirements since 2DLPP involutes
much more coeffcients than classical LPP. A feasible way to deal with this problem
is to use LPP after 2DLPP is applied.
Third, a multilinear generalization of linear discriminant analysis is described.
In some sense, it unifes some existing discriminant analysis methods such as LDA
and 2DLDA. One of important characteristics of multilinear discriminant analysis
is that it can work with the data tensor representation. In addition, it should be
noted that the selection of the optimal dimension of tensors is still an open ques-
tion similar to 2DLDA.
Finally, a lot of experiments are made on face images and handwritten numerical
characters to demonstrate the effectiveness of these tensor subspace methods. It
demonstrates that tensor subspace methods outperform classical subspace methods
in some cases.
r eferences
Chung, F. (1997). Spectral graph theory. In CBMS regional conference series in
mathematics (Vol. 92). RI: American Mathematical Society.
is prohibited.
principal component analysis. IEEE Trans Pattern Anal. Machine Intell., 10(6),
978-983.
Foley, D. H., & Sammon, J. W. (1975). An optimal set of discriminant vector. IEEE
Trans. Computer, C-24(3), 281-289.
He, X. F., & Niyogi, P. (2003). Locality preserving projections. In S. Thrun, L. K.
Saul, & B. Scholkopf (Eds.), Advances in neural information processing systems,16,
153-160. MA: The MIT Press.
Jin, Z., Yang, J-Y., Hu, Z. S., & Luo, Z. (2001). Face recognition based on the uncor-
Jin, Z., Yang, J-Y., Tang, Z., & Hu, Z. (2001). A theorem on the uncorrelated optimal
discriminant vectors. Pattern Recognition, 34(10), 2041-2047.
Jing, X. Y., Zhang, D., & Jin, Z. (2003). Improvements on the uncorrelated optimal
discriminant vectors. Pattern Recognition, 36(8), 1921-1923.
Liang, Z. Z. (2006). High-dimensional discriminant analysis and its application
to color face recognition. Proceedings of the International Conference on Pattern
Recognition (pp. 917-920).
Liang, Z. Z., Shi, P. F., & Zhang, D. (2006). Two-dimensional Fisher discriminant
analysis and its application to face recognition. Proceedings of the 7th Asian Con-
ference on Computer Vision (pp. 130-139).
Liu, K., Cheng, Y-Q., & Yang, J-Y. (1993). Algebraic feature extraction for image
recognition based on an optimal discriminant criterion. Pattern Recognition, 26(6),
903-911.
Schott, J. R. (1997). Matrix analysis for statistics. NY: Wiley-InterScience.
Xiong, H. L., Swam, M. S. S., & Ahmad, M. O. (2005). Two-dimensional FLD for
Xu, Y., Yang, J-Y., & Jin, Z. (2003). Theory analysis on FSLDA and ULDA. Pattern
Recognition, 36(12), 3031-3033.
Xu, Y., Yang, J-Y., & Jin, Z. (2004). A novel method for Fisher discriminant analysis.
Yan, S., Xu, D., Yang, Q., Zhang, L., Tang, X., & Zhang, H. (2005). Discriminant
analysis with tensor representation. Proceedings of the IEEE Conference on Com-
puter Vision and Pattern Recognition (pp. 526-532).
is prohibited.
Yang, J., Yang, J-Y., Frangi, A. F., & Zhang, D. (2003). Uncorrelated projection
discriminant analysis and its application to face image feature extraction. Inter. J.
of Pattern Recognition and Artifcial Intelligence, 17(8), 1325-1347.
Ye, J. P. (2005). Characterization of a family of algorithms for generalized discrimi-
nant analysis on undersampled problems. Journal of Machine Learning Research,
6(Apr), 483-502.
Ye, J. P., Janardan, R., & Li, Q. (2004). Two-dimensional linear discriminant analy-
sis. In Advances in neural information processing systems. San Mateo: Morgan
Kaufmann Publishers.
Ye, J. P., Janardan, R., Park, C. H., & Park, H. (2004). An optimization criterion for
generalized discriminant analysis on undersampled problems. IEEE Trans. Pattern
Anal. Mach. Intell., 26(8), 982-994.
Ye, J. P., & Li, Q. (2005). A two-stage linear discriminant analysis via QR-decom-
position. IEEE Trans. Pattern Anal. Mach. Intell., 27(6), 929-941.

is prohibited.
Chapter X
Tensor Independent
Component Analysis and
Tensor Non-Negative
Factorization
Abstr Act
In this chapter, we describe two tensor-based subspace analysis approaches (tensor
ICA and tensor NMF) that can be used in many felds like face recognition and
other biometric recognition. Section 10.1 gives the background and development
of the two tensor-based subspace analysis approaches. Section 10.2 introduces
tensor independent component analysis. Section 10.3 presents tensor nonnega-
tive factorization. Section 10.4 discusses some potential applications of these two
subspace analysis approaches in biometrics. Finally, we summarize this chapter
in Section 10.5.
10.1 Introduct Ion
Independent component analysis (ICA) (Hyvrinen & Oja, 2001) is a statistical
signal processing technique. The basic idea of ICA is to represent a set of random
variables using basis functions, where the components are statistically independent
or as independent as possible. In general, there are two arguments for using ICA
for image representation and recognition. First, the high-order relationships among
Tensor Independent Component Analysis and Tensor Non-Negative Factorization 203
is prohibited.
images pixels may contain important information for recognition tasks. Second, ICA
seeks to fnd the directions so that the projections of the data into those directions
have maximally non-Gaussian distribution, which may be useful for classifcation
tasks. In addition, the concept of ICA can be viewed as a generalization of PCA,
since it is concerned not only with the second-order dependencies between variables
but also with high-order dependencies between them.
During the past several years, the ICA algorithm has been widely used in face
recognition and biomedical data. Bartlett and Sejnowski (1997) have demonstrated
that the recognition accuracy using ICA basis vectors is higher than that of the PCA
basis vectors with 200 face images. They found that the ICA representation of faces
has the invariance to big changes in pose and small changes in illuminations. In
Bartlett, Movellan and Sejnowski (2002), the authors frst organized the database
into a matrix X where each row vector is a different image. In this representation,
the images are random variables and the pixels are trials. In this case, it makes
sense to talk about independence of images or functions of images. Two images i
and j are independent if when moving across pixels. In addition, they transposed
the matrix X and organized the data so that images are in the columns of X. In this
representation, pixels are random variables and images are trials. Here, it also makes
sense to talk about independence of pixels or functions of pixels. For example, pixel
and would be independent if when moving across the entire set of images. Based
on these two ideas, they suggested two ICA architectures (ICA Architectures I
and II) for face representation and used the Infomax algorithm (Bell & Sejnowski,
1995, 1997) to implement ICA. Both architectures were evaluated on a subset of
the FERET face database and were found to be effective for face recognition. Yuen
and Lai (2000, 2002) adopted the fxed-point algorithm to obtain the independent
components (ICs) and used a householder transform to gain the least square solu-
tion of a face image for representations. Liu and Wechsler (1999, 2003) used an
ICA algorithm to perform ICA and assessed its performance for face identifcation.
All of these researchers claimed that ICA outperforms PCA in face recognition.
Other researchers, however, reported differently. Baek, Draper, Beveridge, and
She (2002) reported that PCA outperforms ICA while Moghaddam (2002), Jin and
Davoine (2004) reported no signifcant performance difference between the two
methods. Socolinsky and Selinger (2002) reported that ICA outperforms PCA on
visible images but PCA outperforms ICA on infrared images.
Recently, Draper, Baek, Bartlett, and Beveridge (2003) tried to account for
these apparently contradictory results. They retested ICA and PCA on the FERET
face database with 1196 individuals and made a comprehensive comparison of the
performances of the two methods and found that the relative performance of ICA
and PCA mainly depends on the ICA architecture and the distance metric. Their
experimental results showed that: (1) ICA Architecture II with the cosine distance
is prohibited.
signifcantly outperforms PCA with L1 (city block), L2 (Euclidean), and cosine
distance metrics. This is consistent with Bartlett and Lius results; (2) PCA with the
L1 distance outperforms ICA Architecture I, which is in favour of Baeks results
(2002); and (3) ICA Architecture II with L2 still signifcantly outperforms PCA with
L2, although the degree of signifcance is not as great as in the ICA Architecture
II with cosine over PCA. Moreover, it should be noted that this latter result is still
inconsistent with Moghaddam (2002) and Jins results (2004). Yang, Zhang, and Yang
(2007) analyzed two ICA architectures and found that ICA Architecture I involves
a vertically centred PCA process (PCA I), while ICA Architecture II involves a
whitened horizontally centred PCA process (PCA II). Thus, it makes sense to use
these two PCA versions as baselines to revaluate the performance of ICA-based
face-recognition systems. Their experiments on the FERET, AR, and AT&T face-
image databases showed no signifcant differences between ICA Architecture I (II)
and PCA I (II), although ICA Architecture I (or II) may, in some cases, signifcantly
outperform standard PCA. It is observed that the performance of ICA strongly de-
pends on the PCA process it involves. Pure ICA projection has only a trivial effect
on the performance in face recognition. More recently, motivated by multilinear
algebra,some researchers proposed tensor independent component analysis.
Non-negative matrix factorization (NMF) (Lee & Seung, 1999, 2001; Li, Hou,
& Zhang, 2001) is also a subspace method, which has been used for image repre-
sentation and image classifcation. Nonnegative matrix factorization differs from
other subspace methods for vector space models, for example, principal compo-
nent analysis (PCA) or vector quantization (VQ), LDA, and ICA due to the use of
constraints that produce nonnegative basis vectors, which make the concept of a
parts-based representation possible. In the case of NMF, the basis vectors contain
no negative entries. As a result, this allows only additive combinations of the vec-
tors to reproduce the original vectors. So the perception of the whole, an image in
a collection, becomes a combination of its parts represented by these basis vectors.
However, some experiments show that directly using the learned feature vectors via
NMF under the Euclidean distance (L2 distance) cannot improve the face recogni-
tion accuracy in contrast to traditional PCA. It is partly because that the learned
bases via NMF are not orthogonal to each other. In order to improve the accuracy,
Li, Hou and Zhang (2001) proposed one method called local NMF which impose
additional constraints on bases. But this method has a slow speed for learning the
bases. Recently Guillamet and Vitria (2003) adopted one relevant metric called earth
movers distance (EMD) for parts-based representation of NMF. However the com-
putation of EMD is too time-demanding. Liu and Zheng (2004) used two strategies
to improve the accuracy of recognition. One is to adopt a Riemannian metric like
distance for the learned feature vectors instead of the Euclidean distance. The other
is to frst orthonormalize the learned bases and then to use the projections of data
is prohibited.
based on the orthonormalized bases for further recognition. Their experiments on
the USPS database demonstrate that the proposed methods can improve accuracy
and even outperform PCA. Up till now, how to use the learned bases and feature
vectors via NMF for further analysis such as recognition is still an open problem.
Note that in the above NMF algorithms the input data must be the vector form.
However, some data are often represented as the tensor form (images). In applying
traditional NMF, the 2D image matrices must be previously transformed into 1D
image vectors. The resulting image vectors usually lead to a high-dimensional image
vector space, where it is diffcult to fnd good bases to approximately reconstruct
original images. That is also the so-called curse of dimensionality problem, which
is more apparent in small-sample-size cases. To this end, some variants of NMF
(Shashua & Hazan, 2005; Zhang, Chen, & Zhou, 2005) are developed in recent
years. The key difference between NMF and its variants is that the latter adopts a
novel representation for original images. Zhang, Chen and Zhou (2005) proposed
2-D non-negative matrix factorization. 2DNMF regards images as 2D matrices and
represents them with a set of 2D bases. Their experimental results on several face
databases show that 2DNMF has better image reconstruction quality than NMF
under the same compression ratio. Moreover, it is also observed that the running
time of 2DNMF is less that of NMF and the recognition accuracy of 2DNMF is
higher than that of NMF.
10.2 t ensor Inde Pendent comPonent AnAl ys Is
10.2.1 t wo Architecture of iCA for image Representations
A. ICA Architecture I
Given a set of M training samples (image column vectors), x
1
,...,x
M
in R
N
, the image
column data matrix is written as X = [x
1
,...,x
M
] and its transpose (image row data
matrix) Y = X
T
.
In Architecture I, the images are seen as random variables and the pixel values
provide observations of these variables. This means that ICA is performed on the
image row data matrix Y. Rewriting Y = [y
1
,...,y
N
], its column vectors y
1
,...,y
N
are
used as observation vectors to estimate the unmixing matrix of the ICA model.
In Architecture I, the following steps are used to extract independent
components.
1. Centering data: Center the data Y in an observation space and obtain its mean
vector
1
{ } (1/ )
N
I j
j
E y N y
=
= =

. Denote
1
[ , , ]
T
I M
= . . Subtracting the
is prohibited.
mean vector from each observation vector, that is,
j j I
y y = - , we obtain the
centered image row data matrix
1
[ , , ]
h N
Y y y = .
2. Sphering data using PCA: Sphere the data using PCA based on the centered
observation vectors
1
, ,
N
y y
. The covariance matrix
I
is given by
1
1
N
T
I i i
i
y y
N
=
=

. (10.1)
Then one can calculate the orthonormal eigenvectors
1
,
...
,
d
(d M) of
I

corresponding to the d largest positive eigenvalues
1
,...,
d
. Letting V =
1
,
...
,
d
and
1
[ , , ]
d
= , one can obtain the whitening matrix
1/ 2
P V
-
= such that
P
T

I
P = I. (10.2)
The data matrix
1
[ , ]
h N
Y y y = can be whitened using the transformation.
R = P
T
Y
h
. (10.3)
3. ICA processing: Perform ICA on R, producing the matrix U
I
with d independent
basis images in its rows,
U
I
= W
I
R
h
, (10.4)
where W
I
is the unmixing matrix generated by a given ICA algorithm based on
the input data R.
Note that the unfxing matrix must be invertible, from Eq. (10.4), it follows
that
1
( )
I I
R W U
-
= . (10.5)
After vertically centering and projecting a given image x in a column vector
onto the row vectors of R, one has z = Rx.
Therefore, in the space spanned by the row vectors of U
I
, i.e., a set of d statistically
independent basis images, the vector of representation coeffcients of image x given
by
1
1
( )
T
s W z
-
=

1
1
( )
T
W Rx
-
= . (10.6)
Overall, the transformation can be obtained by the following two steps:
is prohibited.
z = Rx, (10.7)
1
1
( )
T
s W z
-
= . (10.8)
B. ICA Architecture II
The goal of ICA Architecture II is to fnd statistically independent coeffcients for
the input image data. In this architecture, the images are viewed as observations,
and the pixel values are random variables. ICA is performed directly on the image
column data matrix X. In other words, x
1
,...,x
M
are used as observation vectors to
estimate the unmixing matrix of the ICA model.
Similar to Architecture I, the following steps are used to extract independent
components.
1. Centering data: centralize the data Y in an observation space and obtain its
mean vector
1
{ } (1/ )
M
II j
j
E x M x
=
= =

. When the mean vector is subtracted
from every observation vector, one gets the centered image column data matrix
1
[ , , ]
h M
X x x = .
2. Sphering data using PCA: Sphere the data using PCA based on the centered
observation vectors
1
, ,
M
x x . The covariance matrix
II
is given by
1
1
M
T
II i i
i
x x
M
=
=

. (10.9)
1
,
...
,
d
(d M) of
II

1
,
...
,
d
. Letting V = [
1
,
...
,
d
] and
1
[ , , ]
d
=
, one can obtain the whitening matrix
1/ 2
P V
-
= such that
P
T
I
P = I. (10.10)
The data matrix
1
[ , , ]
h M
X x x = . can be whitened using the transformation
R =P
T
X
h
. (10.11)
3. ICA processing: Perform ICA on R, producing the matrix U
I
with d independent
basis images in its rows, i.e.,
U
I
= W
I
R, (10.12)
is prohibited.
where W
I
the input data R.
that
1
( )
II I
R W U
-
= . (10.13)
I
by
1
( )
T
II
s W z
-
=

1
( )
T T
II
W P x
-
= . (10.14)
Overall, the transform can be obtained by the following two steps:
z = P
T
x, (10.15)
1
( )
T
II
s W z
-
= . (10.16)
From the above analysis, one can see that ICA Architecture I involves a verti-
cally cantered PCA (PCA I), whereas ICA Architecture II involves a whitened
horizontally cantered PCA (PCA II). Standard PCA removes the mean image of
all training samples, while PCA I removes the mean of each image. PCA II is a
whitened version of standard PCA. It can normalize the variances of coeffcients as
well as being able to make these coeffcients uncorrelated. To assess the performance
of two ICA architectures, it is necessary to compare them with the two different
versions of PCA as well as with the standard PCA. In other words, PCA I, PCA II,
and standard PCA should all be used as baseline algorithms to evaluate the ICA.
Figure 10.1 illustrates two ICA-based image-representation architectures.
10.2.2 t he 2dIcA Algorithm
Assume that m n matrices A
k
, k = 1,...,M, denote original M training images. In
2DICA, one does not need to transform the 2D images into its corresponding 1D
vectors. Instead one uses a more straightforward way which views an image as a
2D matrix.
One can use two successive stages to perform 2DICA. First, one aligns the M
images into a m nM matrix X = [A
1
,...,A
M
], where each A
k
denotes one of the M
face images. After X is obtained, we obtain its mean vector
1
(1/ )
nM
j
j
nM X
=
=

and
center the data X. After every vector is subtracted by the mean vector, one can get
is prohibited.
the centered image data matrix
1
[ , , ]
nM
X x x = . One will sphere the data using PCA
based on the centered vectors
1
, ,
nM
x x . The covariance matrix
I
is given by
1
1
nM
T
I i i
i
x x
nM
=
=

. (10.17)
1
,
...
,
d
(d nM) of
II

1
, ,
d
. .
Letting V = [
1
,
...
,
d
] and
1
[ , , ]
d
= , one can obtain the whitening matrix
1/ 2
P V
-
= such that
T
I
P P I = . (10.18)
The data matrix
1
[ , , ]
nM
X x x = can be whitened using the transformation
T
R P X = . (10.19)
Then one can perform ICA on R, producing the matrix U
I
with d independent
basis images in its rows, i.e.,
U
I
= W
I
R, (10.20)
where W
I
the input data R.
that
Figure 10.1. Two ICA-based image-representation architectures
is prohibited.
R = (W
I
)
-1

U
I
. (10.21)
I
by
k
A =

1
( )
T T
I k
W P A
-

T
I k
W P A = k = 1,
...
,M. (10.22)
It is obvious that
k
A
(k = 1,
...
,M) are d n matrices.
Based on Eq. (10.22), we perform the second stage on
k
A (k =1,...,M). We
frst construct a matrix from submatrices
k
A (k =1,...,M) and defne as follows
1

[ , , ]
T T
M
Y A A = . It is obvious that Y is a n dMmatrix. After Y is obtained, we
obtain its mean vector
1
(1/ )
dM
j
j
dM Y
=
=

and center the data Y. After the mean
vector is subtracted from every vector, we get the centered image data matrix
1
[ , , ]
dM
Y y y =
. We will sphere the data using PCA based on the centered vectors
1
, ,
dM
y y
. The covariance matrix
II
is given by
1
1
dM
T
II i i
i
y y
dM
=
=

. (10.23)
Then we can calculate the orthonormal eigenvectors
1
,...,
d
(d dM) of
II

1
,...,
d
. Letting V =
1
,...,
d
and
1
[ , , ]
d
=
, we obtain the whitening matrix
1/ 2
P V
-
=
such that
T
II
P P I =

. (10.24)
The data matrix
1
[ , , ]
dM
Y y y = can be whitened using the transformation
T
R P Y =

. (10.25)
Then we perform ICA on R, producing the matrix U
I
with d independent basis
images in its rows, i.e.,
I II
U W R = , (10.26)
where
II
W is the unmixing matrix generated by a given ICA algorithm based on
the input data R.
that
is prohibited.
1
( )
II I
R W U
-
= . (10.27)
I
by
k
A =
```

1
( ) ( )
T T T
II k
W P A
-

( ) ( )
T T T
II k I
W P A P W =

, k = 1,...,M. (10.28)
It is obvious that
k
A
```
are d d matrices.
From the above analysis, we know that 2DICA frst performs the column ICA,
then followed by row ICA operations. Figure 10.2 demonstrate how a high-dimen-
sional image is converted into a low-dimensional image by using 2DICA.
Finally, the pseudo-code for performing 2DICA is descried as follows:
The 2DICA Algorithm
Input: m n image matrices
1
{ }
M
k k
A
=
and the reduction dimension d.
Output: the matrices W
I
, ,
II
W P P
.
Step1: Align the M images into a m nM matrix X = [A
1
,
...
,A
M
]
Step2: Perform classical ICA for X, and obtain matrices W
I
and P.
Step3: Construct a new matrix
1

[ , , ]
T T
M
Y A A =
Step4: Performing classical ICA on Y, and obtain matrices
II
W and P
,
Step5: Obtain the approximate matrix of
k
A
```

( ) ( )
T T T
II k I
W P A P W =

k = 1,
...
,M.
Figure 10.2. The dimensionality reduction by using 2DICA
k
A Column
ICA
k
A

k
A

I
C
A

R
o
w

T T
k II k I
A W P A PW =
```
is prohibited.
Note that in the above algorithm, we frst align the M images into a m nM
matrix X = [A
1
,
...
,A
M
]. In fact, we can also align the M images into a m nM matrix
1
[ , , ]
T T
M
Y A A = .
10.2.3 multilinear IcA
Multilinear ICA is obtained by frst representing the data tensor G as the node-n
product of N mode matrices
G = S
1
U
1
...
N
U
N
.
Then the N-mode ICA algorithm mainly consists of the following two steps:
1. For n = 1,
...
,N, compute the mode matrix U
n
.
2. obtain the core tensor as follows:
1 1
1 1 N N
S G U U
- -
=
.
As stated in Section 10.2.1, there are also two strategies for performing multi-
linear independent components analysis (MICA). Architecture I forms a factorial
code, where each set of coeffcients that encodes people, viewpoints, illuminations,
etc., is statistically independent, while Architecture II is to fnd a set of independent
bases across people, viewpoints, illuminations, etc.
Architecture I: Transforming the fattened data tensor G into the nth mode and
computing the ICA in classical mode, we can get
( )
T T T
n n n n
G V U =
,
1
( )( )
T T
n n n n
V W WU
-
=
,
T T
n n
K U = , (10.29)
where G
(n)
is the mode-n fattened of G . In other words, we perform the singular
value decomposition on G
(n)
.
The columns associated with each of the mode matrices, U
n
, are statistically
independent. The relationship between N-mode ICA and N-mode SVD in the context
of Architecture I can be represented as follows:
1 1 N N
G Z U U =
is prohibited.
1 1 1 1
1 1 1
1 1 1 1
1 1
( )
T T T T
N N N N
T T
N N N
T T
N N N N
N N
Z UW W U W W
Z UW U W
Z W W U U
C U U
- -
- -
- -
=
=
=
=
. (10.30)
Note that Z is the core tensor in N-mode SVD.
Architecture II: Flattening the data tensor G in the nth mode and computing
the ICA, we obtain:
( )
T
n n n n
G U V =
,
1
( )( )
T
n n n n
U W W V
-
= ,
= U
n
K
n
, (10.31)
where G
(n)
is the mode-n fattened of G. In other words, one performs the singular
value decomposition on G
(n)
.
Architecture II results in a set of basis vectors that are statistically independent
across the different modes. The relationship between N-mode ICA and N-mode
SVD in the context of Architecture I is as follows:
1 1
1 1
1 1 1 1
1 1
1 1 1
1 1
1 1 1 1
1 1
( )
N N
N N N N
N N N
N N N N
N N
G Z U U
Z UWW U W W
Z UW U W
Z W W U U
C U U
- -
- -
- -
=
=
=
=
=
. (10.32)
Note that Z is the core tensor in N-mode SVD.
10.3 t ensor non- neg At Ive f Act or IzAt Ion (nf )
10.3.1 classical nmf
The key characteristic of NMF is the non-negativity constraints imposed on the two
factors, and the non-negativity constraints are compatible with the intuitive notion
of combining parts to form a whole. NMF can also be interpreted as a parts-based
is prohibited.
representation of the data due to the fact that only additive, not subtractive, com-
binations are allowed. Assume that the image database is represented as an N M
matrix V, each column containing N non-negative pixel values corresponding to
one of the M face images. In order to compress data or reduce the dimensionality,
NMF fnds two nonnegative matrix factors W and H such that
1
( ) ( )
d
i iu ia a
a
V WH W H
=
=
(10.33)
Here the d columns of W are called NMF bases, and the columns of H are its
combining coeffcients. The dimensions of W and H are N d and d m respec-
tively. The rank (d) of the factorization is usually chosen such that (N+M)r< NM,
and hence the compression or dimensionality reduction is achieved. The compres-
sion ratio of NMF is easily gotten as (Nd+Md).
To obtain an approximate factorization V W H, a cost function is needed to
quantify the quality of the approximation. NMF uses the following divergence
measure as the objective function:
1 1
( , ) ( ln ( ) )
( )
N M
ij
ij ij ij
i j
ij
V
D V WH V V WH
WH
= =
= - +
, (10.34)
After some simplifcation and elimination of pure data terms, one can obtain
1 1 1 1
( , ) ( ln )
N M d d
ik kj ij ik kj
i j k k
D V WH W H V W H
= = = =
= -

. (10.35)
Taking the derivative with respect to H gives
1 1
1
( , )
N N
ib ia
ia d
i i
ab
ik kb
k
V W
D V WH W
H
W H
= =
=
= -
. (10.36)
Then the gradient algorithms can be described as follows:
ab ab ab
H H -
( , )
ab
D V WH
H
, (10.37)
ab ab ab
H H -
1 1
1
( )
N N
ib ia
ia d
i i
ik kb
k
V W
W
W H
= =
=
-

, (10.38)
where
ab
is a step size.
is prohibited.
Let
ab
=.
1
ab
N
ia
i
H
W
=
Then the multiplicative rule is:

1 1
1
/
N d
ia ib ik kb
i k
ab ab N
ia
i
W V W H
H H
W
= =
=
. (10.39)
Similarly, taking the derivative with respect to W gives
1 1
1
( , )
M M
cj dj
dj d
j j
cd
ck kj
k
V W
D V WH H
H
W H
= =
=
= -
. (10.40)
Then the gradient algorithms can be described as follows:
cd cd cd
H H -
( , )
cd
D V WH
H
, (10.41)
cd cd cd
H H -

1 1
1
( )
M N
cj dj
dj d
i j
ck k
k
V W
H
W H
= =
=
-

, (10.42)
where v
cd
is a step size.
Let v
cd
=
1
cd
M
dj
j
W
H
=
.Then the multiplicative rule is :

1 1
1
/
M d
di ci ck ki
i k
cd cd M
di
i
H V W H
W W
H
= =
=
. (10.43)
Overall, the pseudo-code for computing the matrices W and H is described as
follows:
t he Algorithm of NMF
Input: the N M matrix V, and reduction dimension d.
Output: Matrices W and H
Step1:initialize the matrices W and H and set k 1.
is prohibited.
Step2: while not convergent
(a) update the bases W using Eq. (10.43)
(b) update the coeffcients H using Eq. (10.42)
Step3: end while
10.3.2 t he 2dnmf Algorithm
k
, k = 1,...,M, denote original M training images. In
2DNMF, one does not need to transform the 2D images into its corresponding 1D
vectors. Instead, one can use a more straightforward way which views an image
as a 2D matrix.
In (Zhang, Chen, & Zhou, 2005), The authors used two successive stages to
perform 2DNMF. First, they align the M images into a m nM matrix X = [A
1
,...,A
M
],
where each A
k
denotes one of the M face images. After X is obtained, one can perform
classical NMF on the matrix X. That is, one needs to fnd a m d non-negative
matrix L and a d nM non-negative matrix H such that
X LH , (10.44)
Note that L and H are the bases and combining coeffcients respectively. Parti-
tioning the matrix H into M submatrices, denoted as H = [H
1
,
...
,H
M
], one can write
Eq. (10.44) as follows
1 1
[ , , ] [ , , ]
M M
A A L H H . (10.45)
Furthermore, one can obtain
A
k
= LH
k
(k = 1,...,M). (10.46)
From Eq. (10.46), one can note that the image A
k
is a weighted sum of the column
bases L. Since H
k
is an d n matrix, there are dn coeffcients for the image A
k
. In
order to further reduce the number of coeffcients, Zhang, Chen, and Zhou (2005)
performed the second stage to fnd row bases. They constructed a matrix from
submatrices [H
1
,
...
,H
M
] and defned as follows
1
[ , , ]
T T
M
H H H = . It is obvious that
H is a nd M matrix. Based on
H, they seeked a n d non-negative matrix R and

a d dM non-negative matrix C such that
H RC . (10.47)
is prohibited.
Here R and C are the bases and combing coeffcients respectively. It is obvi-
ous that R and C can be obtained by performing classical NMF on the matrix
H
. Partitioning the matrix C into M submatrices, denoted as C = [C
1
,
...
,C
M
], one can
write Eq. (10.47) as follows
1 1 1
[ , , ] [ , , ]
T T
M
H C R C C . (10.48)
Furthermore, one can obtain
T
k k
H RC (k =1,
...
, M). (10.49)
From Eq. (10.49), one can note that the image
T
k
H
is a weighted sum of the rows
bases.
Substituting Eq. (10.49) into Eq. (10.46), one can obtain
T T
k k
A LC R (k =1,
...
, M). (10.50)
From Eq. (10.50), one can fnd that each image is decomposed into three matrices.
Note that C
k
is a coeffcient matrix and its dimension is d d.
Let L = [l
1
,
...
,l
d
] and R = [r
1
,
...
,r
d
]. Defne the outer product between the column
base l
i
and the row base r
j
as follows.
ij i j
E l r = 1 i d, 1 j d, (10.51)
According to Eq. (10.51), one can write Eq. (10.50) into
T T
k k
A LC R

1 1
( )
d d
k ij ij
i j
D E
= =
=
. (10.52)
From Eq. (10.52), the image A
k
is a linear combination of 2D bases. Note that
2D bases E
ij
have the following properties.
(a) the dimension of E
ij
is equal to the size of the original images.
(b) The intrinsic dimensionality of E
ij
is 1.
From the above analysis, one can know that 2DNMF frst performs the column
NMF, then followed by row NMF operations. Figure 10.3 shows how a high-dimen-
sional image is transformed into a low-dimensional image by using 2DNMF.
Finally, the pseudo-code for performing 2DNMF is descried as follows.
is prohibited.
t he 2dnmf Algorithm
1
{ }
M
k k
A
=
and the reduction dimension d.
Output: the matrices L, R, and D
k
(k = 1,
...
,d).
Step1: Align the M images into a m nM matrix X = [A
1
,...,A
M
]
Step2: Perform classical NMF on X = LH, and obtain matrices L and H.
Step3: Partition H into H = [H
1
,...,H
M
], construct a new matrix from H, denoted
as
1
[ ]
T T
M
H H H =
Step4: Performing classical NMF on
H RC = , and obtain matrices R and C,

Step5: Obtain the approximate matrix of
T T
k k
A LC R (k = 1,...M).
Note that in the above algorithm, one frst aligns the M images into a m nM
matrix X = [A
1
,...,A
M
]. In fact, one can also align the M images into a n mM matrix
1
[ , , ]
T T
M
Y A A = . The dimension d in two stages may be different. It can be bigger
than or smaller than the dimension of images.
10.3.3 multilinear non-negative f actorization
A tensor G of dimensions [d
1
]
...
[d
n
] is indexed by n indices i
1
,
...
,i
n
with 1 i
j
d
j
. It
is of rank at most k if it can be expressed as a sum of k rank-1 tensors, i.e. a sum
of n-fold outer-product:
1 2
1
k
j j j
n
i
G u u u
=
=
. Although the rank of a matrix

can be found in the polynomial time using the SVD algorithm, the rank of a tensor
is an NP-hard problem.
Given an n-way array G, we expect to fnd a non-negative rank-k tensor
1 2
1
k
j j j
n
i
u u u
=

described by nk vectors
j
i
u
. The following optimization
problem is constructed:
Figure 10.3. An illustration of dimensionality reduction by using 2DNMF
is prohibited.
2
1
1
min ( ) min
2
k
j j
i i
i
F
f u G u
=
= -
, (10.53)
subject to, 0,
j
i
u
where F is the square Frobenious norm, i.e., the sum of squares of all entries of
the tensor elements. The direct approach is to form a positive preserving gradient
descent scheme.
Let , A B denote the inner-product operation, i.e.
1 1
1
, ,
, , n n
n
i i i i
i i
A B

. Since the
differential commutes with inner-products, i.e., , 2 , d A A A dA = , we have
1 1
1 1
1
,
2
k k
n j n j
i i i i
j j
d G u G u
= =
= =
- -

,
1 1
1 1
, [ ]
k k
n j n j
i i i i
j j
G u d G u
= =
= =
- -

. (10.54)
Taking the differential with respect to
s
r
u and noting that
1
1
[ ]
k
n j
i i
j
d G u
=
=
-

1
1 1
( )
r s s n s
i i r i r i
u d u u
-
= = +
= -
,
The differential becomes
1
1 1 1
1
( ) , ( )
k
s n j r s s n s
r i i i i r i r i
j
df u u u d u u
-
= = = +
=
=

1
1 1
, ( )
r s s n s
i i r i r i
G u d u u
-
= = +
- .
The differential with respect to the lth coordinate
s
rl
u is
1
1 1 1
1
( ) ,
k
s n j r s n s
rl i i i i l i r i
j
df u u u e u
-
= = = +
=
=

1
1 1
,
r s n s
i i l i r i
G u e u
-
= = +
- ,
where e
l
is the lth column of the d
r
d
r
identity matrix. Let
1
[ ] [ ]
n
S d d denote
an n-tuple index {i
1
,
...
,i
n
}. Let /
r
S i denote the set
1 1 1
{ , , , , , }
r r n
i i i i
- +
and
r
i l
S

denote
the set of indices S where the index i
r
is replaced by the constant l. Then, using the
identity
1 1 2 2
, x y x y , one can obtain the partial derivative
,
1 /
( ) ( )
i m
r
r
k
j j T s s
rl i i S l m i
s
j S i i r m r
rl
f
u u u G u
u

=
= -

. (10.55)
is prohibited.
Using the multiplicative update rule by setting the constant
s
rl
of the gradient
descent formula
s
rl
u
s s
rl rl
s
rl
f
u
u
to be,
s
rl

1
( )
s
rl
k
j j T s
rl i i
j i r
u
u u u
=
=

, (10.56)
one can obtain
s
rl
u

,
/
1
( )
i m
r r
s s
rl S l m i
S i m r
k
j j T s
rl i i
j i r
u G u
u u u

. (10.57)
It is shown that this update rule can make the value of the optimization function
become small. The following proposition will provide the theoretical foundation
for the update rule.
Proposition10.1: Let f(x
1
,...,x
n
) be a real quadratic function with Hessian of the
form H = cI with c > 0. Given a point
1
( , , )
t t t
n
x x x =
and a point
1
( ( )
t t t
x x f x
+
= -
with
1
0
c
< <
, then f(x
t+1
) < f(x
t
).
In this section, the performances of 2DICA, ICA, 2DNMF, NMF and tensor NF
are demonstrated. Zhang, Chen, and Zhou (2005) computed the bases of NMF and
2DNMF from the training set. They also computed the bases of PCA and 2DPCA
[26]. They did the experiment on the FERET face databases as shown in Section
3.4.4, and 200 face images are used as the training set. Figure 10.4 plots parts of
the bases gotten from the four methods respectively. The bases in Figure 10.4 (a)
and (b) are obtained by using the method (Zhang, Chen, & Liu, 2005). For NMF
the frst 16 columns of W are transformed into the representation in matrix form,
whereas for 2DNMF the 16 bases (1i4, 1j4) are directly described as the im-
ages. From Figure 10.4, one can see that both the bases of PCA and 2DPCA are
global. It is also observed that 2DPCA possess some strips or blocks like structures.
A possible explanation is that 2DPCA is essentially a kind of line-based PCA. As a
result, the bases of 2DPCA cannot yet refect any local or part-based features. On the
other hand, Figure 10.4 (c) indicates that although NMF is a part-based algorithm,
is prohibited.
its bases still take on some holistic properties similar to PCA. In contrast, one can
notice from Figure 10.4 (d) that the bases of 2DNMF are much sparser than those
of NMF. It is worth noting that although the bases of 2DNMF are sparse, they have
no parts-based (like eye, mouth, etc. in face image) features any more due to the
essence of 2D methods. That is, 2DNMF is essentially a kind of line-based NMF.
Consequently, what 2DNMF really learns is some parts of lines. Moreover, because
each base of 2DNMF can be generated using a d-dimensional column base and
a d-dimensional row base, its storing coeffcients (d+d) are much fewer than that
of the NMF base (which is d*d). Thus one can use much more sparse-distributed
2D bases to represent the original image. As a result, in some sense, 2DNMF has
higher compression ratio than NMF.
The 1th row of Figure 10.5 shows some of the NMF factors generated from a
set of 2429, 1919, images from the MIT CBCL database. One can clearly see that
ghost structures and the part decomposition is complicated .The NTF factors (rank-1
matrices), shown in the 2nd row, have a sharper decomposition into sparse compo-
nents. The 2nd row shows an overlap of rank-1 factors whose energy is localized in
the same image region. One can clearly see that the parts (which now correspond to
higher rank matrices) corresponding to eyes, cheeks, and shoulders are shown.
Figure 10.4. (Zhang, Chen, & Zhou, 2005). Bases obtained from PCA (a), 2DPCA
(b), NMF (c) and 2DNMF (d) respectively. For NMF the frst 16 columns of matrix
W are retransformed to matrix for plotting, while for 2DNMF the 16 bases Eij
(1i4, 1j4) are directly plotted as images.
is prohibited.
In the following, we compare the performances of NMF, 2DNMF, ICA and
2DICA when they are used for the classifcation tasks. The recognition accuracy,
which is defned as the percentage of correctly recognized images in test images,
is used as the performance measure. The experiment is frst made to classify test
images on two face databases: UMIST and Yale Face images. The UMIST Face
Database consists of 564 images of 20 people. Each covers a range of poses from
profle to frontal views. Subjects cover a range of race/sex/appearance. Each subject
exists in their own directory and images are numbered consecutively as they were
taken. Each image is resized to 112 92 pixels. As for the description of the Yale
face database, please refer to Section 3.3.3. In this database, the facial portion of
each original image was automatically cropped based on the location of eyes and
the cropped image was resized to 8080 pixels. In both face databases, we randomly
choose fve images from each person as training samples and others are used as test-
ing samples. In addition, the experiments are made to test the classifcation accuracy
of the fourth methods under partial occlusions (10%). In order to reduce variations,
the fnal results are averaged over ten runs. Table 10.1 demonstrate the experimental
results. Note that the dimensions are chosen based on the Fisher criterion. The 3rd
and 4th rows denote performance under partial occlusions. From Table 1, one can
observe that the 2D Methods (including 2DICA and 2DNMF) outperform classical
1D methods. Moreover, NMF is not superior to ICA in terms of the classifcation
performance. In the case of occlusions, the performances of two methods become
worse. These experiments show that it is reasonable to adopt the 2D methods for
feature extraction in real-world applications.
Figure 10.5. (Shashua & Hazan, 2005). The frst row: leading NMF factors of CBCL
face dataset, compared to leading NTF factors in the 2nd row. The 3rd row: summed
factors of NTF located in the same region (resulting in higher rank factors) see
texts for further explanation.
is prohibited.
10.5 summAr y
In this chapter, two tensor subspace methods: tensor ICA and tensor NMF, are
introduced. Although independent components analysis (ICA) maximizes the
statistical independence of the representational components of a training image
ensemble, it cannot distinguish between the different factors, or modes, inherent
to image formation including scene structure, illumination, and imaging. Fortu-
nately, the multifactor models for ICA can deal with these problems. Specifcally,
the multilinear ICA (MICA) model of image ensembles can learn the statistically
independent components of multiple factors. In addition, one can notice that ICA
employs linear (matrix) algebra, whereas MICA exploits multilinear (tensor) algebra.
Different from ICA, NMF is an effective method for fnding parts-based represen-
tation of non-negative data such as face images. Although it has been successfully
applied in real world, directly using NMF for face recognition often leads to low
classifcation performance. Moreover, when performing on large databases, NMF
needs considerable computational costs. To overcome the disadvantages of classical
NMF, tensor NMF is developed. The key difference between classical NMF and
tensor NF and is that the latter adopts a novel representation for original images.
Specifcally, NMF is a special case of tensor NF based on multilinear algebra.
r eferences
Baek, K., Draper, B. A., Beveridge, J. R., & She, K. (2002). PCA vs. ICA: A com-
parison on the FERET data set. Proc. Joint Conf. Inf. Sci. (pp. 824-827).
Bartlett, M. S., Movellan, J. R., & Sejnowski, T. J. (2002). Face recognition by
independent component analysis. IEEE Trans. Neural Netw., 13(6), 1450-1464.
Bartlett, M. S., & Sejnowski, T. J. (1997). Independent of face images: a representation
for face recognition. Proceedings of the Fourth Annual Joint Symposium on Neural
computation, 3299, 528-539.
ICA 2DICA NMF 2DNMF
UMIST 73.45(5.31) 75.68(6.30) 71.22(7.22) 72.34(4.33)
Yale 70.13(8.37) 72.37(7.25) 66.90(7.39) 68.76(9.12)
UMIST(10%) 68.39(6.34) 69.68(8.15) 60.05(8.99) 62.33(9.44)
Yale(10%) 61.22(7.15) 69.23(10.12) 59.22(8.34) 64.67(9.56)
Table 10.1. Performance comparisons of four methods on two face databases (%)
is prohibited.
Bell, A. J., & Sejnowski, T. J. (1995). An information-maximization approach to
blind separation and blind deconvolution. Neural Comput., 7(6), 1129-1159.
Bell, A. J., & Sejnowski, T. J. (1997). The independent components of natural
scenes are edge flters, Vis. Res., 37(23), 3327-3338.
Draper, B. A., Baek, K., Bartlett, M. S., & Beveridge, J. R. (2003). Recognizing
faces with PCA and ICA. Comput. Vis. Image Underst., 91(1-2), 115-137.
Guillamet, D., & Vitria, J. (2003). Evaluation of distance metrics for recognition
based on non-negative matrix factorization. Pattern Recognition Letters, 24(9-10),
1599-1605.
Hyvrinen, A., & Oja, E. (2001). Independent component analysis: algorithms and
applications. Neural Netw., 13(4-5), 411-430.
Jin, Z., & Davoine, F. (2004). Orthogonal ICA representation of images. Proc. 8th
Int. Conf. Control, Autom., Robot. and Vis. (pp. 369-374).
Lee, D. D., & Seung, H. S. (1999). Learning the parts of objects by non-negative
matrix factorization. Nature, 401, 788-791.
Lee, D. D., & Seung, H. S. (2001). Algorithms for non-negative matrix factorization.
Adv. Neural Inform. Process. Syst., 13, 556-562.
Li, S. Z., Hou, X. W., & Zhang, H. J. (2001). Learning spatially localized parts-
based representation. Proceedings of IEEE International Conference on Computer
Vision and Pattern Recognition (pp. 207-212).
Liu, C., & Wechsler, H. (1999). Comparative assessment of independent component
analysis for face recognition. Proc. 2nd Int. Conf. Audio and Video-Based Biometric
Person Authentication (pp. 211-216).
Liu, C., & Wechsler, H. (2003). Independent component analysis of Gabor features
for face recognition. IEEE Trans. Neural Netw., 14(4), 919-928.
Liu, W. X., & Zheng, N. N. (2004). Non-negative matrix factorization based methods
for object recognition. Pattern Recognition Letters, 25(8), 893-897.
Moghaddam, B. (2002). Principal manifolds and probabilistic subspaces for visual
recognition.IEEE Trans. Pattern Anal. Mach. Intell., 24(6), 780-788.
Shashua, A., & Hazan, T. (2005). Non-negative tensor factorization with application
to statistics and computer vision. Proceedings of the Intl Conference on Machine
Learning (pp. 792-799).
is prohibited.
Socolinsky, D., & Selinger, A. (2002). A comparative analysis of face recognition
performance with visible and thermal infrared imagery. Proc. Int. Conf. Pattern
Recog. (pp. 217-222).
Yang, J., Zhang, D., & Yang, J-Y. (2007). Constructing PCA baseline algorithms to
reevaluate ICA-based face-recognition performance. IEEE Trans. Systems, Man,
and CyberneticsPart B: Cybernetics, 37(4), 1015-1021.
Yuen, P. C., & Lai, J. H. (2000). Independent component analysis of face images.
Proc. IEEE Int. Conf. Biological Motivated Comput. Vis., 545-553.
Yuen, P. C., & Lai, J. H. (2002). Face representation using independent component
analysis. Pattern Recognit., 35(6), 1247-1257.
Zhang, D. Q., Chen, S. C., & Liu, J. (2005). Representing image matrices: Eigen-
images vs. Eigenvectors. Proceedings of the International Symposium on Neural
Networks (ISNN05), vol. 2, Chongqing, China (pp. 659-664).
Zhang, D. Q., Chen, S. C., & Zhou, Z. H. (2005). Two-dimensional non-negative
matrix factorization for face representation and recognition. Proceedings of the
International Conference on Computer Vision Workshop on Analysis and Modeling
of Faces and Gestures (pp. 35.-363)
is prohibited.
Chapter XI
Other Tensor Analysis and
Further Direction
Abstr Act
In this chapter, we describe tensor-based classifers, tensor canonical correlation
analysis and tensor partial least squares, which can be used in biometrics. Section 11.1
gives background and devolvement of these tensor methods. Section 11.2 introduces
tensor-based classifers. Section 11.3 gives tensor canonical correlation analysis
and tensor partial least squares. We summarize this chapter in Section 11.4.
11.1 Introduct Ion
In general, a biometric system consists of data acquisition phase, feature extraction
phase and the classifcation phase. In the data acquisition phase, the data obtained
are often represented by multidimensional arrays, that is, tensors, such as the grey
face image, the colour faces in image classifcation and the gene expression data. In
the feature extraction phase, multilinear subspace methods mentioned in Chapters
VIII, IX, and X can be used for data representation and feature extraction. In the
classifcation phase, the classifers (Bousquet, Boucheron, & Lugosi, 2004; Duda,
Hart, & Stock, 2001; Muller, Mika, Ratsch, Tsuda, & Scholkopf, 2001; Highleyman,
1962) play an important role in the biometric system and how to design a good
classifer is of interest for researchers. Traditionally, the classifer design is almost
based on vector pattern, that is, before using them, any non-vector pattern such as
an image should be frst vectorized into the vector pattern by the techniques such
Other Tensor Analysis and Further Direction 227
is prohibited.
as concatenation. However, the Ugly Duckling Theorem (Chen, Wang, & Tian,
2007) indicates that it cannot be said that one pattern representation is always bet-
ter than another. As a result, it is not always reasonable to design classifers based
on traditional vector patterns.
Motivated by the tensor ideas in feature extraction, some researchers have designed
some classifers based on tensor ideas in recent years. For example, Chen,Wang and
Tian (2007) designed the classifers using a set of given matrix patterns. They frst
represented a pattern in matrix form and extended existing vector-based classif-
ers to their corresponding matrixized versions. Specifcally, considering a similar
principle to the support vector machine which maximizes the separation margin
and has superior generalization performance, the modifed HK algorithm (MHKS)
(Leski, 2003) is chosen and then a matrix-based MHKS (MatMHKS) classifer is
developed. Their experimental results on ORL, Letters and UCI data sets show
that MatMHKS is more powerful in generalization than MHKS. Further, Wang
and Chen (2007) proposed a new classifer based on matrix patterns and LS-SVM.
This method is referred to as MatLSSVM. The MatLSSVM method can not only
directly operate on original matrix patterns, but also effciently reduce memory for
the weight vector in LS-SVM. However, one of disadvantages of MatLSSVM is that
there exist unclassifable regions when it is extended to the multi-class problems.
To avoid this point, a corresponding fuzzy version of MatLSSVM (MatFLSSVM)
is further proposed to remove unclassifable regions for multi-class problems. Ex-
perimental results on some benchmark datasets show that their proposed method
is competitive in classifcation performance compared to LS-SVM and fuzzy LS-
SVM (FLS-SVM). In Tao, Li, Hu, Maybank and Wu (2005; Tao, Li, Wu, Hu, &
Maybank, 2006) a supervised tensor learning (STL) framework is established for
convex optimization techniques such as support vector machines (SVM) and mini-
max probability machines (MPM). Within the STL framework, many conventional
learning machines can be generalized to take nth-order tensors as inputs. These
generalized algorithms have several advantages: (1) reduce the problem of the
curse of dimensionality in machine learning and data mining; (2) avoid the failure
to converge; and (3) achieve better separation between the different categories of
samples. Further, they generalized MPM to the STL version, which is called tensor
MPM (TMPM). The TMPM method can obtain a series of tensor projection vector
by an iterative algorithm. The experiments on a binary classifcation problem show
that TMPM signifcantly outperforms the original MPM.
In addition to introducing tensor-based classifers, we also demonstrate more
examples of tensor analysis including tensor canonical correlation analysis and tensor
partial least squares. Canonical Correlation Analysis (CCA), frstly developed by
Hotelling (1936) in 1930s, is a powerful statistical technique to measure the linear
relationship between two multidimensional variables. It fnds two linear combina-
is prohibited.
tions, one for each multidimensional variable, which are optimal with respect to
correlations. In other words, the goal of CCA is to fnd the directions of maximum
data correlation. Up till now, the CCA method has been widely used in several
felds such as signal processing, medical studies, and pattern recognition (Melzer,
Reiter, & Bischof, 2003; Sun, Zeng, Liu, Heng, & Xia, 2005; Zheng, Zhou, Zou, &
Zhao, 2006). However, this method always suffers from the small sample size (SSS)
problem in the applications. To this end, two-dimensional CCA based on classical
CCA is developed. Sample images in 2DCCA are represented as 2D matrices instead
of vectors, and discriminative features are extraction from the image covariance
matrix. In addition, partial least squares (PLS) (Wold, 1982, 1985; Helland, 1988;
Oskuldsson, 1988; Lewi, 1995) are developed for the regression problem and have
proven to be popular and effective approaches to the problems in biometrics such
as predicting the bioactivity of molecules. The PLS method can be used for multi-
variate as well as univariate regressions. Moreover, the PLS algorithms can reduce
the variance by introducing a small bias. Since Wold (1982, 1985) proposed PLS in
the feld of econometrics, the PLS methods had found a variety of applications in
the areas such as bioinformatics, face recognition and character recognition. For
example, in (Tan, Shi, Tong, Hwang, & Wang, 2004), the authors used discriminant
partial least squares (DPLS) to classify different types of human tumors and showed
good prediction performance of DPLS on four microarray datasets. In Baek and
Kim (2004) the PLS algorithm is used to extract facial features and the experimental
results show the PLS algorithm can obtain lower recognition error rate than PCA.
However, classical PLS often suffers from the small sample size problem. To this
end, 2DPLS and multilinear PLS are developed in this chapter.
11.2 t ensor- bAsed cl Ass If Iers
11.2.1 t ensor ho-Kashyap (hK) Classifer
11.2.1.1 Classical Ho-Kashyap and MHKS Algorithm
For a two-class
1 2
( , ) problem, assume that {( , ) 1, , } { 1, 1}
N
i i
S x y i M = = + -
is a set of M samples. Let [ ,1]
T T
i i
x x = . In general, a linear discriminant function for
two-class problems can be defned as
( )
T
i i
g x u x = , (11.1)
is prohibited.
where u is the weight vector or a projection vector. If ( )
i
g x is bigger than zero,
the corresponding sample
i
x
is considered as
1
class. Otherwise, it belongs to
2

class.
Further, applying the class label, one can obtain
( ) 0
T
i i i i
y g x y u x = > , (i = 1,
...
,M). (11.2)
Eq. (11.2) shows that the samples can be correctly classifed by the linear
discriminant function ( )
i
g x .
Let
1 1
[ , , ]
T
M M
Y y x y x = . We denote Eq. (11.2) in matrix form
0 Yu >
. (11.3)
In (Leski, 2003), the optimization function of the Ho-Kashyap (HK) algorithm
is defned as the following quadratic loss function:
2
( , ) J u b Yu b = - , (11.4)
where b is the margin vector and b > 0.
Taking the derivative of Eq.(11.4) with respect to u and b gives
2 ( )
T
J
Y Yu b
u
= -
, (11.5)
2( )
J
Yu b
b
= - -
. (11.6)
In addition, for any value of b, one has
u Y b
+
= , (11.7)
where Y
+
denotes the pseudo-inverse.
Note that b > 0 and b should avoid converging to zero in the gradient descent
procedure. Consequently, the following update rule for b is adopted.
(1) 0
( 1) ( ) ( ( ) | ( ) |)
b
b k b k e k e k
>
+ = + +
, (11.8)
is prohibited.
where ( ) ( ) ( ) e k Yu k b k = - , k is the index of iterations and is the learning rate. To
avoid the HK algorithm being sensitive to outliers, Leski (2003) modifed the HK
algorithm and defned the canonical hyperplane as
1
1
M
Yu

. (11.9)
Based on Eq.(11.9), Leski (2003) developed the following optimization function
to obtain the projection vector u and the margin vector b,
2
1
min ( , ) min( 1 )
T
N
I u b Yu b Cu u
= - - + , (11.10)
where C is the regularization parameter. The optimal parameters u and b are obtained
by alternately optimizing
1
1
( ) ( 1 )
T T
M
u Y Y CI Y b
-
= + + , (11.11)
1
1
M
b Yu

= - , (11.12)
where I is a matrix by setting the last element of an identity matrix to zero and 1
M1

denotes a column vector whose elements are 1.
11.2.1.2 Matrix-Pattern-Oriented HK Classifer
k
, k = 1,
...
,M, denote M images and y
k
, k = 1,
...
,M, are
the corresponding M labels.
Chen et al. (2007) developed a matrix-pattern-oriented, two-sided linear clas-
sifer based on the HK algorithm with the regularization learning, as an extension
to the HK algorithm.
Based on the matrix notation, they defned the following decision function
( )
T
f A u Av = + b
0
, (11.13)
where
m
u ,
n
v , and b
0
is a bias.
From Eq. (11.13), one can design a classifer based on the matrix pattern. In such
a case, the two-sided linear function for two-class
1 2
( , ) problems can be given
0
( )
T
i i
g A u Av b = +

1
2
0
0
i
i
A
A
>
<
. (11.14)
is prohibited.
Further, from Eq. (11.14), one has
0
( ) ( ) 0
T
i i i i
y g A y u Av b = + > , (i = 1,
...
,M). (11.15)
Similar to classical HK algorithm, a canonical hyperplane is defned as
0
( ) ( ) 1
T
i i i i
y g A y u Av b = + >
, (i = 1,
...
,M). (11.16)
Based on Eq. (11.16), Chen et al. (2007) defned the following optimization
problem
min I (u, v, b
0
, b)
2
0 1 2
1
min[ ( ) 1 ) ( )]
M
T T T
i i i
i
y u Av b b C u S u v S v
=
= + - - + +
,
(11.17)
where b = [b
1
,...,b
M
]
T
, S
1
= mI and S
2
= nI are two regularization matrices, the
regularization parameter C controls the generalization ability of the classifer.
Let
1
[ , , ]
T
M
Y y y = ,
[ ,1]
T T
i i i
y y u A =
, and
0
[ , ]
T T
v v b = . We transform Eq. (11.17)
into the following matrix form
2
1 1 2
min ( , , ) min(|| 1 || ( ))
T T
N
I u v b Yv b C u S u v S v
= - - + + , (11.18)
where
2
S is a matrix with dimension of (n + 1) (n + 1) and

2
S
2
0
0 0
S
(
=
(

.
Taking the derivative of the function with respect to , u v and b gives
0 1
1
2 [ ( ) 1 ]
M
T
i i i i N i
i
I
y Av y u Av b b
u

=
= + - -

+
1
2CS u, (11.19)
1
2( 1 )
M
I
Yv b
b

= - - -
, (11.20)
1
2( 1 )
M
I
Yv b
v

= - - -
. (11.21)
By setting 0
I
u
and 0
I
v
, one can get

1
2 1
( ) (1 )
T T
M
v Y Y CS Y b
-
= + + , (11.22)
is prohibited.
1
1 0
1 1
( ) ( (1 ) )
M M
T T
i i i i i i
i i
u Avv A CS y b y b Av
-
= =
= + + -

. (11.23)
From Eqs. (11.22) and (11.23), one can notice that the weight vectors u and v
are
all determined by the margin vector b whose components determine the distance
of the corresponding sample to the separation hyperplane. Moreover, because u and
v are mutually dependent, it is necessary to use the iterative procedure to obtain
them.
Based on the above discussion, Chen et al. (2007) summarized MatMHKS as
follows.
The algorithm for matrix-pattern-oriented HK classifers
Step1: Fix 0, 0 1 C < < , initialize b(1) 0 and u(1), set k = 1
Step2:
1
2 1
( ) (1 ( ))
T T
M
v Y Y CS Y b k
-
= + +
Step3:
1
1
M
e Yv b
= - -
Step4: ( 1) ( ) ( ( ) | ( ) |, b k b k e k e k + = + +
Step5: if || ( 1) ( ) || b k b k + - > , then k = k + 1, goto Step (6); else stop
Step 6:
1
1 0
1 1
( 1) ( ( ) ( ) ) ( (1 ( ) ) ( ))
M M
T T
i i i i i i
i i
u k Av k v k A CS y b k y b Av k
-
= =
+ = + + -

, go to
step 2
where k denotes the index of iterations, and is a preset parameter.
Note that this algorithm is degenerated into MHKS when m=1, u=1 and step 6
is omitted. When C is set to 0, the algorithm is degenerated into the original Ho-
Kashyap algorithm. As a result, Ho-Kashyap and MHKS algorithms are two special
cases of MatMHKS. In addition, Chen et al. (2007) pointed out that there exists the
relationship between MatMHKS and MHKS by the Kronecker product. That is,
MatMHKS is equivalent to MHKS with the Kronecker product decomposability
constraint. In some sense, MatMHKS is guided by some prior information such as
the structural or local contextual information which is refected in the representa-
tion of the Kronecker product. Moreover, the advantages of MatMHKS are that it
can avoid overtraining and reduction in dimensionality for weights.
11.2.2 t ensor ls -svm
11.2.2.1 Classical LS-SVM
Let {( , ) 1, , } { 1, 1}
N
i i
S x y i M = = + - be a set of examples. In LS-SVM
(Suykens & Vandewalle, 1999), the decision function is written as
is prohibited.
1
( ) ( , )
M
i i i
i
f x y K x x b
=
= +
, (11.24)
where
i
are a real constant , b is a bias, and K(x, x
i
) is a kernel function .The output
of f(x) is 1 if its value is greater than 0, -1 otherwise. In general, parameters b and
i
can be obtained by solving the following primal optimization problem
2
1
1
min
2 2
i
M
T
i
C
w w
=
+

, (11.25)
subject to ( ( ) ) 1 , 1, ,
T
i i i
y w x b i M + = - = . .
where () is a linear or nonlinear function which maps the input space into a higher
dimensional feature space, w is a weight vector to be determined, C is a regulariza-
tion constant and
i
is a slack variable.
From Eq. (11.25), one can construct the following Lagrangian function
2
1 1
1
( , , , ) [ ( ( ) ) 1 ]
2 2
M M
T T
i i i i i i i
i i
C
L w b w w y w x b
= =
= + - + - +

.
(11.26)
Taking the derivative with respect to , , ,
i i
w b gives
1
0 ( )
M
i i i
i
L
w y x
w
=
= =

, (11.27)
1
0 0
M
i i
i
L
y
b
=
= =

, (11.28)
0
i i
i
L
C
= =
, (11.29)
0 0 ( ( ) ) 1 0
T
i i i
i
L
y w
x b
=
= +
- + =
. (11.30)
From Eqs. (11.27), (11.28), (11.29) and (11.30), one can obtain
is prohibited.
1
0 0
1
T
b Y
a Y C I
-
( - ( (
=
( ( (
+

(11.31)
where Y = [y
1
,...,y
M
]
T
, ()
ij
= y
i
y
j
K(x
i
, x
j
) and a = [
1
,
...
,
M
].
From Eq. (11.31), one can obtain a and b. Thus, a decision function in Eq.(11.24)
is obtained.
11.2.2.2 Matrix LS-SVM
k
, k =1,
...
,M denote M images and y
k
, k =1,
...
,M are the
corresponding label k =1,
...
,M.
Based on LS-SVM, Wang and Chen (2007) proposed a variant of LS-SVM,
called MatLSSVM, for two-class problems. This method can directly deal with
the data in the form of matrices. In their method, the decision function is defned
as the following form:
( )
T
f A u Av = + b, (11.32)
where u R
m
and v R
n
.
It is obvious that the MatLS-SVMs method needs to store m + n coeffcients
for u and v, while the classical LSSVMs method needs to store mn coeffcients. As
a result, the higher dimension the pattern has, the less memory the MatLSSVMs
method needs to store weight vectors.
In MatLSSVMs, the following optimization problem is constructed:
2
1
1 1
min
2 2 2
i
M
T T
i
C
u u v v
=
+ +
. (11.33)
subject to ( ) 1
T
i i i
y u Av b + = - . i = 1,
...
,M.
From the above optimization problem, one can see that it is diffcult to fnd the
optimal u and v simultaneously. However, when one fxes one of variables (u,v),
it is easy to solve Eq. (11.33). In other words, the variable u can be obtained by a
similar procedure of LS-SVMs in the case of the fxed v and the variable v can be
obtained by a similar procedure of LS-SVMs in the case of the fxed u.
Wang and Chen (2007) do not use this alternating optimization procedure to
obtain u and v. They obtain u and v by the following equations:
1
M
i i i
i
u y Av
=
=
, (11.34)
is prohibited.
1
1
( )
M
T
t t i i i t
i
v v y A u
+
=
= - -
, (11.35)
where is the learning rate and t is the iterative index.
Overall, the algorithm of MatLSSVMs proposed by Wang and Chen (2007) is
summarized as follows.
The Algorithm of MatLSSVMs
Step1: initialize v
0
, , MaxIter, , and C.
Step2: For t = 1: MaxIter
(a) obtain
i
by solving classical LSSVM in the fxed v
t-1
and obtain
1
1
M
t i i i t
i
u y Av
-
=
=
by
update v by Eq.(11.35),
(b) if
1 t t
v v
+
- > and t < MatIter, then step 2, else go to step 3.
Step3: stop
From the algorithm of MatLSSVMs, one can see that the MatLSSVMs method
consists of two phases. First, the solution to v can be gotten by an iterative algo-
rithm. Second, for every fxed v, an analytical solution to u similar to LS-SVM can
be obtained. Here, for a fxed v, the solution to u is free of local minima due to the
fact that it is a convex optimization problem. However, in solving v, its optimality
may not be guaranteed due to non-convexity of the criterion for v. In addition, it is
worth noting that if we want the problem to be convex, we may alternatively take
only u instead of both u and v as unknown and view the vector v as a hyperparam-
eter in such a case.
11.2.3. Classifcation Learning Based on Convex Optimization
Programming
11.2.3.1 Convex Programming
In general, learning models can be formulated as optimization problems. In this
subsection, we frst introduce convex optimization and then give a general formula-
tion for convex optimization-based learning.
is prohibited.
The convex optimization problem has the following expression
min f
0
(u) (11.36)
subject to ( ) 0(1 )
i
f u i m ,
( ) 0( 1 )
i
f u m i p = + ,
where f
i
(u) (0 i m) are convex functions and
( )( 1 )
i
f u m i p +
are affne
functions.
The convex optimization problem defned in Eq.(11.36) consists of a large number
of popular special cases, such as the linear programming (LP) (Vanderbei, 2001), the
linear fractional programming (LFP) (Boyd & Vandenberghe, 2004), the quadratic
programming (QP), the quadratically constrained quadratic programming (QCQP),
the second-order cone programming (SOCP) (Lobo, Vandenberghe, Boyd, & Lebret,
1998), the semidefnite programming (SDP) (Vandenberghe & Boyd, 1996), and the
geometric programming (GP) (Boyd, Kim, Vandenberghe, & Hassibi, 2006). All of
these special cases have been widely applied in different areas, such as computer
networks, machine learning, computer vision, psychology, the health research, the
automation research, and economics.
The signifcance of a convex optimization problem is that the solution is unique
(i.e., the locally optimal solution is also the globally optimal solution). Consequently,
the convex optimization theory has been widely applied to pattern classifcation
for many years, such as LP in the linear programming machine (LPM) (Hardoon,
Szedmak, & Taylor, 2004) QP in the support vector machines (SVM), SDP in the
distance metric learning (DML).
In the following, we give a general formula for convex optimization-based learn-
ing
min f(u,b,
i
) (11.37)
subject to y
i
c
i
(u
T
x
i
+ b)
i
(i = 1,
...
,M),where c
i
: R
N+M+1
R, f: R
N+M+1
R a criterion
(convex function) for classifcation0, c
i
: R
N+M+1
R for all 1 i M are convex
constraint functions, x
i
1 i M are training samples and their class labels are
given by y
i
R
N
{1,-1},
1
,
...
,
M
are slack variables, and u R
N
and b R determine
the classifcation hyperplane, that is, sign(u
T
x+b). By defning different classifcation
criteria f and convex constraint functions c
i
, we can obtain a large number of learning
machines, such as SVM, MPM, and DML.
11.2.3.2 Supervised Tensor Learning: A Framework
Supervised tensor learning (STL) (Tao et al., 2005) generalizes the vector-based
learning algorithms to accept general tensors as input. In STL, assume that there are
M training samples
1 N
I I
i
A

(i = 1,
...
,M)represented by tensors associated with
is prohibited.
class label information
{1, 1}
i
y -
.One wants to separate the positive sample (y
i
= 1
) from the negative samples (y
i
= 1) based on a criterion. This extension is obtained
by using
1 N
I I
i
A

. Therefore, STL is defned by
1 1
min ({ } , ,{ } )
N M
k k i i
f u b
= =
, (11.38)
subject to
1
( )
N
i i i k k i
k
y c A u b
=
+
, 1 i M.
There are two different points between the vector-based learning and the ten-
sor-based learning: (1) the training samples are represented by vectors in vector-
based learning, whereas they are represented by tensors in tensor-based learning;
and (2) the classifcation decision function which is defned by u R
N
and b
R

in vector-based learning is sign(u
T
x + b), whereas the classifcation deci-
sion function which is defned by u
i
R
N
and b R in tensor-based learning is
1
( )
N
k k k
sign A u b
=
+ . In the vector-based learning, one has the classifcation hyper-
plane, (u
T
x + b) = 0, whereas in the tensor-based learning, one defnes the classifcation
tensorplane, that is,
1
( ) 0
N
k k k
A u b
=
+ = .
From Eq.(11.38), one can obtain the following Lagrangian function
1 1 1 1
1 1
1 1
1 1 1
({ } , ,{ } , ) ({ } , ,{ } ) ( ( ) )
({ } , ,{ } ) ( )
N M
N M N M
k k i i k k i i i i i i k k i
i k
N M M
N M
k k i i i i i i k k i i
i i k
L u b f u b y c A u b
f u b y c A u b
= = = =
= =
= =
= = =
= - + -
= - + +

(11.39)
where
1
,
...
,
N
are Lagrangian multiplies.
Further, the solution of Eq. (11.39) can be obtained by solving the following
saddle point of the optimization function.
1 1
({ } , ,{ } , )
N M
k k i i
L u b
= =
. (11.40)
The derivative of
1 1
({ } , ,{ } , )
N M
k k i i
L u b
= =
with respect to u
j
is

(11.41)
1 1
1 1
1
( ( ))
( )
( )
N M
i i i i k k
i k
j j j
N M
i
i i i k k
i k
j j
M
i
i i i j j
i
j j
L f
y c A u b
u u u
dc f
y A u b
u dz u
dc f
y A u
u dz u
= =
= =
=

= - +

= - +

= -

,
is prohibited.
where
1
N
i k k
k
z A u b
=
= +
.
The derivative of
1 1
({ } , ,{ } , )
N M
k k i i
L u b
= =
with respect to b is
L
b

f
b

1 1
( ( ))
N M
i i i i k k
i k
y c A u b
b
= =

=
f
b

1 1
( )
N M
i
i i i k k
i k
dc
y A u b
dz b
= =

=
f
b

1
M
i
i i
i
dc
y
dz
=
, (11.42)
where
1
N
i k k
k
z A u b
=
= +
.
To obtain a solution of STL, one needs to set
L
b
0 and 0
j
f
u
. Further, one
has
,
1
( )
M
i
i i i j j
i
j j
dc f
y A u
u dz u
=

=

(11.43)
f
b

1
M
i
i i
i
dc
y
dz
=
, (11.44)
Based on Eq. (11.43), one can notice that the solution u
j
depends on u
k
(1 k
N, k j). That is, directly obtaining the solution of STL is intractable. Fortunately,
the alternating projection scheme can provide a strategy for obtaining the solution
of STL. The key idea in the alternating projection optimization for STL is to obtain
u
j
with the fxed u
k
(1 k N, k j)in an iterative way.
Finally, Tao et al. (2005) proposed the following algorithm for the supervised
tensor learning.
The Algorithm for the Supervised Tensor Learning
Input: The training samples
1 N
I I
i
A

and class label information y
i

{1,1}
i = 1,
...
,M.
Output: The parameters in classifcation tensor plane
1
{ }
N
k k
u
=
and b, such that the
STL objective function
1 1
({ } , ,{ } )
N M
k k i i
f u b
= =
defned in Eq. (11.38) is minimized.
Step1: Set
1
{ }
N
k k
u
=
equal to random unit vectors,
is prohibited.
Step2: Carry out steps 35 iteratively until convergence;
Step3: For k = 1 to N:
Step4: Obtain
1
{ }
N
k k
u
=
by optimizing
Step5: end
Step6: convergence checking. if the calculated
1
{ }
N
k k
u
=
have been converged. Here
u
i,t
is the current projection vector and u
i,t-1
is the previous projection vector.
11.3 other t ensor subs PAce AnAl ys Is
11.3.1 t ensor Canonical Correlation Analysis
11.3.1.1 classical cc A
Given a set of samples S = (x
1
, y
1
),...,(x
M
, y
M
) of (x, y), where x
i
, y
i
R
N
. Let X =
[x
1
,
...
, x
M
] and Y = [y
1
,
...
, y
M
]. Choosing a direction u R
N
and projecting X onto the
direction u, we have
T
X X u = . (11.45)
Choosing a direction v R
N
and projecting Y onto the direction v, one has
T
Y Y v = . (11.46)
The frst stage of canonical correlation is to choose u and v such that the correlation
between two vectors is maximized. In other words, the following objective function
needs to be maximized
=

max ( , ) corr X Y
=

, ,
max max
T T
T T
X Y X u Y v
X u Y v
X Y
< > < >
=
, (11.47)
where <
.
,
.
> denotes the Euclidean inner product of the vector and is the norm
of a vector. If we use Ef (x,y) to denote the empirical expectation of the function
f(x,y), then
1
1
( , ) ( , )
M
i i
i
Ef x y f x y
M
=
=

. (11.48)
is prohibited.
Based on Eq. (11,48), we can rewrite Eq. (11.47) as
( ) ( )
max max
( ) ( ) ( ) ( )
T T T T
T T T T
E u xy v u E xy v
E x u E y v E x u E y v
= =
. (11.49)
Note that the covariance matrix of (x,y) is
( , ) [
T
xx xy
yx yy
C C x x
C x y E
C C y y
(
( | || |
= = (
( | |
( \ .\ .

. (11.50)
The total covariance matrix C(x,y) is a block matrix where the within-set covari-
ance matrices are C
xx
and C
yy
, and the between-set covariance matrices are C
xy
and
C
yx
.Using Eq. (11.50), we can rewrite the function as

max
T
xy
T T
xx yy
u C v
u C u v C v
=
. (11.51)
Note that the solution of Eq. (11.51) is not affected by scaling u and v either
together or independently. In other words, replacing u by u gives the same results
for .Based on this point, the following conditions must be imposed in order to
maximize the function .
1
T
xx
u C u = , (11.52)
1
T
yy
v C v = . (11.53)
From Eqs. (11.51), (11.52) and (11.53), one can obtain the corresponding
Lagrangian
( , , ) ( 1) / 2
T T
xy x xx
L u v u C v u C u = - -

( 1) / 2
T
y yy
v C v - -
. (11.54)
Taking derivatives with respect to u and v, one can obtain
0
xy x xx
L
C v C u
u
= - =
, (11.55)
0
yx y yy
L
C u C v
v
= - =
. (11.56)
From Eqs. (11.55) and (11.56), one further has
0
T
y yy
v C v = -,
T
x xx
u C u
(11.57)
is prohibited.
which together with the constraints implies that
y x
= . Let
y x
= = . Assume
that C
yy
is invertible, one has
1
/
yy yx
v C C u
-
= . (11.58)
Substituting Eq. (11.58) into (11.55) gives
1 2
xy yy yx xx
C C C u C u
-
= . (11.59)
It is obvious that Eq. (11.59) is a generalized eigenproblem. Hence, we can
obtain projection vectors by solving this generalized eigenproblem. In some real
applications, some regularization parameters are used to avoid the singularity
problem in the covariance matrices.
11.3.1.2 2dcc A
k
, k = 1,
...
, M, denote M images and m n matrices B
k
,
k = 1,
...
,M , denote another M images. In 2DCCA, one does not need to transform
the 2D images into its corresponding 1D vectors. Instead one uses a more straight-
forward way which views an image as a matrix.
We use two successive stages to perform 2DCCA. First, we align the M images
(A
k
) into a m nM matrix X = [A
1
,
...
,A
M
] and align the M images (B
k
) into a m nM
matrix Y = [B
1
,
...
,B
M
]. After X and Y is obtained, one can perform classical CCA on
the matrices X and Y. That is, one needs to fnd two directions u, v R
m
such that
the following objective function is maximized:
,
max
T T
T T
X u Y v
X u Y v
< >
= (11.60)
In such a case, we obtain u and v by performing classical CCA on Eq.(11.60).
After obtaining u and v, one can easily compute
T
X X u = , (11.61)
T
Y Y v = . (11.62)
It is straightforward to verify that
X and
Y are nM 1 vectors. Further, we parti-

tion vectors
X and
Y into M blocks respectively, each block having dimensions of n,

denoted by
1

( ) [( ) , , ( )]
T T
M
X X X = and
1

( ) [( ) , , ( )]
T T
M
Y Y Y = . Let
1

[ , , ]
M
X X X =
is prohibited.
and
1

[ , , ]
M
Y Y Y = . Note that X and Y are n M matrices. In the second phage, we
continue to perform classical CCA on the matrices X and Y. That is, one needs to fnd
two directions u , v R
n
such that the following objective function is maximized:
,
max

T T
T T
X u Y v
X u Y v
< >
=
(11.63)
Likewise, it is not diffcult to obtain u and v by solving Eq.(11.63).
From the above analysis, we briefy summary the 2DCCA algorithm as
follows.
t he 2dcc A Algorithm
1
{ }
M
k k
A
=
and
1
{ }
M
k k
B
=
.
Output: the vectors u and v, u and v
Step1: Align the M images
1
{ }
M
k k
A
=
into a m nM matrix X = [A
1
,
...
,A
M
] and Align
the M images
1
{ }
M
k k
B
=
into a m nM matrix Y = [B
1
,
...
,B
M
],
Step2: Perform classical CCA on X and Y, and obtain vectors u and v,
Step 3: Compute
T
X X u = and
T
Y Y v = and par t it ion
X and
Y i nto
1

( ) [( ) , , ( )]
T T
M
X X X = and
1

( ) [( ) , , ( )]
T T
M
Y Y Y = . Let
1

[ , , ]
M
X X X = and
1

[ , , ]
M
Y Y Y = .
Step4: Performing classical CCA on X and Y, and obtain vectors u and v,
Step5: Compute
T
k k
A u A u
T
k k
B v B v k = 1,
...
, M).
Note that in the above algorithm, we frst align the M images into a m nM
matrix X = [A
1
,
...
,A
M
]. In fact, we can also align the M images into a n mM matrix
1

[ , , ]
M
Y Y Y = . In addition, we only obtain two pairs of vectors (u,v) and ( u, v). In real
applications, it is often needs multiple pairs of vectors. The methods for obtaining
multiple pairs of vectors in 2DCCA are similar to those in classical CCA.
11.3.1.3 Multilinear CCA
Assume that A
k

1 N
I I

and B
k
1 N
I I

, k = 1,..., M, are two groups of tensors.
Given
1 1
(1) I R
U

,
2 2
(2)
,
I R
U

,
( )
N N
I R N
U

, and
1 1
(1) I R
V

,
2 2
(2)
,
I R
V

,
( )
N N
I R N
V

, we consider the following spaces
(1) ( ) N
U U and
(1) ( ) N
V V
. The tensors , A
k
, k = 1,
...
, M, are projected onto
(1) ( ) N
U U
, denoted by
k
A
(1) ( )
1
(( ) ( )
T N T
k N
A U U =
and the tensors B
k
, k = 1,
...
, M, are projected onto
(1) ( ) N
V V
, denoted by
k
B

(1) ( )
1
(( ) ( )
T N T
k N
B V V =
.
Similar to classical CCA, multilinear CCA is to fnd the optimal projection
matrices
1 1
(1) I R
U

,
2 2
(2)
,
I R
U

,
( )
N N
I R N
U

such that the following objective
function is maximized.
is prohibited.

,
max

( ) ( )
E A B
E A E B
< >
=
, (11.64)
where
1
1

( , ) ,
M
i i
i
E A B A B
M
=
= < >
,
1
1

( ) || ||
M
i
i
E A A
M
=
=

and
1
1

( ) || ||
M
i
i
E B B
M
=
=

.
It seems that it is very diffcult for us to deal with Eq. (11.64). However, the
matrix representation of tensors provides an effective strategy for simplifying Eq.
(11.64). Note that two tensors satisfying the following relationship
(1) (2) ( )
1 2
( ) ( ) ( )
T T N T
N
S A U U U = . (11.65)
can be represented in the following matrix form:
( ) ( 1) ( ) (1) ( 1)
( ) ( )
( ) ( )
n T n N n
n n
S U A U U U U
+ -
= . (11.66)
It is not diffcult to verify that
2
2
( ) n
F
F
A A = , for 1, , n N = . . For the sake
of notational simplicity, let
( 1) ( ) (1) ( 1) (\ )
( )
n N n n
U U U U U
+ -
=
and
( 1) ( ) (1) ( 1) (\ )
( )
n N n n
V V V V V
+ -
= . In such a case, we have
( ) ( )

( ) , ( ) ,
i n i n i i
A B A B < >=< >=<
( ) (\ )
( )
( ) ( ) ,
n T n
i n
U A U
( ) (\ )
( )
( ) ( )
n T n
i n
V B V >
= trace [
( ) (\ )
( )
( ) ( )
n T n
i n
U A U

(\ ) ( )
( )
( ) ( ) ( )
n T T n
i n
V B V
] (11.67)
2
( ) ( )
1 1 1
( ) (\ ) ( ) (\ )
( ) ( )
1 1
( ) (\ ) (\ ) ( )
( ) ( )
1 1

( ) , ( )
( ) ( ) , ( ) ( )
[( ) ( ) ( ( ) ) ]
M M M
i i n i n
i i i
M M
n T n n T n
i n i n
i i
M M
n T n n T n
i n i n
i i
A A A
U A U U A U
trace U A U A U U
= = =
= =
= =
=< >
=< >
=

. (11.68)
In a similar way, we have
2
1
M
i
i
B
=
= trace [
( ) (\ )
( )
1
( ) ( )
M
n T n
i n
i
V B V
=

(\ ) ( )
( )
1
( ( ) )
M
n T n
i n
i
B V V
=
]. (11.69)
is prohibited.
In general, in order to obtain the optimal projections U
(1)
, U
(2)
,
...
,U
(N)
, it is necessary
to maximize Eq.(11.64). We note that it is very diffcult to fnd the optimal projections
simultaneously. To this end, an iterative algorithm is developed for obtaining the
optimal projections U
(1)
, U
(2)
,
...
,U
(N)
, like the algorithm of HOSVD. Further, we can
see that for fxed U
(1)
,
...
, U
(n - 1)
,
...
,U
(n + 1)
,
...
, U
(N)
it is not diffcult to obtain the optimal
projection U
(n)
similar to the CCA algorithm.
Let
( ) n
AA
C = [
(\ )
( )
1
( )
M
n
i n
i
A U
=

(\ )
( )
1
( ( ) )
M
n T
i n
i
A U
=
],
( ) n
BB
C =
[
(\ )
( )
1
( )
M
n
i n
i
B V
=

(\ )
( )
1
( ( ) )
M
n T
i n
i
B V
=
],
and
( ) n
AB
C =[
(\ )
( )
1
( )
M
n
i n
i
A U
=

(\ )
( )
1
( ( ) )
M
n T
i n
i
B V
=
].
Assume that
( ) n
BB
C is invertible. Then the optimal projection U
(n)
can be obtained
by solving the generalized eigenvalue problem
( ) ( ) 1 ( ) ( ) ( ) ( )
( )
n n n n n n
AB BB BA AA
C C C U C U
-
= . (11.70)
and V
(n)
can be obtained by
V
(n)
=
1
( ) 1 ( ) ( )
2
( ) .
n n n
BB BA
C C U
-
-
(11.71)
As a summary of the above discussion, the algorithm is shown as follows.
The iterative algorithm of Multilinear CCA
In:
1 N
I I
i
A

,
1 N
I I
i
B

for i = 1,
...
, M
Out: U
(n)
, V
(n)
n = 1,
...
, N.
Step1: Initial values:
( )
0
n
U and
( )
0
n
V (2 n N);
Step2: Compute the mean tensor M
i
of class i and the globle mean tensor M;
Step3: iterative until convergence
(a) (i) compute
(2) ( ) (2) ( )
1 1
( ), ( )
N N
k k k k
E U U F V V = =

1 (1) 1 (1) 1
1
( ) [( ) ][( ) ]
c
T
b i i i
i
S n M M E M M E
=
= - -
,
(ii) obtain
(1)
1 k U +
by solving Eq.(11.70) and
(1)
1 k
V
+
by Eq.(11.71)
(N) (i) compute
(1) ( 1) (1) ( 1)
1
( ), ( )
N N
N k k k k
E U U F V V
- -
= = ,

(ii) obtain
(1)
1 k U + by solving Eq.(11.70) and
(1)
1 k
V
+
by Eq.(11.71)
is prohibited.
Obtain converged values
(1) ( ) (1) ( )
, , , , ,
N N
U U V V . .
.
Step4: Compute
(1) ( )
1
( ) ( ) ( )
T N T
i i N
A A U U = ,

(1) ( )
1
( ) ( ) ( )
T N T
i i N
B B V V = for i = 1,
...
, M.
From the above algorithm, we note the following facts: when A
i
(i = 1,
...
, M)
are vectors, the algorithm is degenerated into classical CCA. Therefore, in some
sense, the proposed method unifes current CCA and 2DCCA. It is obvious that
multilinear CCA can directly deal with multidimensional data. Since the algorithm
is iterative, the solution is theoretically local optimal. Like the HOSVD algorithm,
probably the algorithm has several optimal solutions. In addition, we can see the
algorithm involves setting initial values and judging the convergence. Therefore,
it is necessary to further discuss these two problems in the future.
11.3.2 t ensor Partial l east squares
11.3.2.1 Classical Partial Least Squares
Let X= [x1,
...
,xM]
T
be the (M N) matrix of M input and Y be M L matrix of the
corresponding L-dimensional response, where x
i
(i = 1,
...
,n) R
N
. Without loss of
generality, we assume that the matrix X and Y should be centralized in the PLS
regression. They are easily obtained by a translation of data. In general, PLS is to
approximate X and Y by some variables and to model the relationship between X
and Y. Specifcally, the PLS method decomposes X and Y matrices into the follow-
ing form:
T
X TP E = +
(11.72)
T
Y UQ F = + (11.73)
where T and U are M k matrices of the extracted k score vectors , P and Q are
matrices of loadings, and E and F are matrices of resides.
In addition, the PLS regression model can also be represented with the regression
coeffcient matrix B and the residual matrix R as follows:
Y XB R = + (11.74)
1
( )
T T
B W P W C
-
= , (11.75)
is prohibited.
where P is the N k matrix containing loading vectors. The matrix P
T
W is upper
triangular and invertible. Note that Rannar, Lindgram, Geladi, and World (1994)
obtained the following equalities:
1
( )
T T
P X T T T
-
= , (11.76)
1
( )
T T
C Y T T T
-
= , (11.77)
T
W X U =
, (11.78)
T = XW
1
( )
T
P W
-
, (11.79)
Substituting Eqs. (11.76), (11.77) and (11.78) into Eq. (11.75) and using the
orthogonality of the matrix T, we obtain B in the following form:
1
( )
T T T T
B X U T XX U T Y
-
= . (11.80)
Further, substituting into Eqs. (11.78) and (11.79) into Eq. (11.80), one can
obtain
1
( )
T T T T
B W W X XW W X Y
-
= . (11.81)
It is obvious that Eq.(11.81) only contains an unknown matrix W.
Assume that Y can be constructed by XB when enough dimensions are adopted.
In other words, R is a zero matrix. In such a case, we obtain
Y = X
1
( )
T T T T
W W X XW W X Y
-
. (11.82)
We multiple Y
T
= XW in right-side in Eq. (11.82) and multiple X
T
in the left-side
in Eq. (11.82). Then we obtain
T T T
X YY XW X X =

1
( )
T T T T T
W W X XW W X YY XW
-
. (11.83)
Let
1
( )
T T T T T
W X XW W X YY XW
-
= . Then we obtain
T T T
X YY XW X XW = . (11.84)
It is not diffcult to verify that is a diagonal matrix when W consists of the
eigenvector matrix of Eq. (11.84) in Fukunaga (1990).
is prohibited.
Note that the solutions of Eq. (11.84) are conjugate orthogonal with respect to
X
T
X. However, in some PLS regression problems, it is required the W should be a
column orthogonal matrix. In such a case, we can obtain the orthogonal matrix W
by Gram- Schmidt orthogonalization. In addition, we can see that X
T
X is singular
when the number of samples is smaller than the dimension of samples. Fortunately,
some effective algorithms (Howland & Park, 2004; Ye, Janardan, Park, & Park,
2004) for overcoming the singular problem in generalized eigenvalue problems
have been developed recently.
11.3.2.2 2d Partial l east squares (2dPls )
k
, k = 1,..., M, denote M images and
Y is an M c
matrix of the corresponding c-dimensional response. In 2DPLS, one does not need
to transform the 2D images into its corresponding 1D vectors.
First, we align the M images (A
k
) into a mM n matrix
1
[ , , ]
T T T
M
X A A = .
Let
1 1

[ , , , , , , ]
M M
m m
Y Y Y Y Y =
_ _
,
where
i
Y is the transpose of the ith row of matrix
i
Y. After X and Y is obtained,
one can perform classical PLS on the matrices X and Y. That is, one needs to fnd
projection matrices W
1
R
n d
such that
1 1
T T T
X YY XW X X W = . (11.85)
It is obvious that Eq. (11.85) is a generalized eigenvalue problem. It is not diffcult
to verify that is a diagonal matrix when
W
1
consists of the eigenvector matrix of Eq. (11.85).
Further, we obtain the following matrix:
1
1 1 1 1
( )
T T T T
B W W X XW W X Y
-
= . (11.86)
Second, we align the M images (A
k
) into a nM m matrix U = [A
1
,...,A
M
]
T
. Let
1 1

[ , , , , , , ]
M M
n n
T Y Y Y Y =
_ _
,
where
i
Y is the ith column of matrix
T
Y . After U and T are obtained, one can perform
classical PLS on the matrices U and T. That is, one needs to fnd the projection
matrices W
2
R
m d
such that
is prohibited.
2 2
T T T
U TT UW U U W = . (11.87)
It is obvious that Eq. (11.87) is a generalized eigenvalue problem. It is not diffcult
to verify that is a diagonal matrix when W
2
consists of the eigenvector matrix of
Eq. (11.87).
1
2 2 2 2
( )
T T T T
C W W U UW W U T
-
= . (11.88)
From Eqs. (11.86) and (11.88), one can notice that B is an n c matrix and C is
an m c matrix. After a test sample A is put, we have
T
Y C AB =
(11.89)
From
Y, one can see that the response of each sample is a vector. However, from
Eq. (11.89), we can know Y is a c c matrix. To be consistent with the original form,
one can obtain the response of a test sample by using Y, denoted by
( )
final
Y mean Y =
, or max(Y), or min( ) Y . (11.90)
From the above analysis, we briefy summary the 2DPLS algorithm as
follows.
The 2DPLS Algorithm
1
{ }
M
k k
A
=
and the responses
1
{ }
M
k k
Y
=
.
Output: the vectors B and C
Step1: Align the M images (A
k
) into a nM m matrix
1
[ , , ]
T T T
M
X A A = and Let
1 1

[ , , , , , , ]
M M
m m
Y Y Y Y Y =
_ _
, where
i
Y is the transpose of the ith row of matrix
Y.
Step 2: Perform classical PLS B on X and Y, obtain W
1
by Eq. (11.85) and B by
Eq.(11.86)
Step3: align the M images (A
k
) into a nM m matrix U = [A
1
,...A
M
]
T
. Let
1 1

[ , , , , , , ]
M M
n n
T Y Y Y Y =
_ _
,
where
i
T
Y .
Step4: Performing classical PLS on U and T, and obtain W
2
by Eq. (11.87) and C
by Eq. (11.88),
Step5: compute the fnal response by Eq. (11.90).
is prohibited.
11.3.2.3 Multilinear Partial Least Squares
Assume that A
k
1 N
I I

, k = 1,
...
,M,are M tensors and
Y is a M c matrix of the
corresponding c-dimensional response. The matrix representation of tensors pro-
vides an effective strategy for dealing with multilinear PLS.
Let B be an
1 N
I I M tensor which consists of a group of tensors A
k
1 N
I I

and B
(n)
be an
1 1 1
( )
n n N n
I I I I M I
- +

matrix by unfolding the matrix B. Let
1 1 1 1 1 1
( ) 1 1

[ , , , , , , ]
n n N n n N
n M M
I I I I I I I I
Y Y Y Y Y
- + - +
=

_ _
,
where
i
T
Y .
After B
(n)
and Y
(n)
are obtained, one can perform classical PLS on the matrices B
(n)

and Y
(n)
. That is, one needs to fnd the projection matrices W
(n)

n d
I R
such that
( ) ( )
( ) ( ) ( ) ( ) ( ) ( )
T T n T n
n n n n n n
B Y Y B W B B W =
. (11.91)
( ) ( ) ( ) ( ) 1 ( )
( ) ( ) ( ) ( )
(( ) ) ( )
n n n T T n n T T
n n n n
C W W B B W W B Y
-
= . (11.92)
After a test sample A is put, we have
(1) (2) ( )
1 2
( ) ( ) ( )
T T N T
N
Y A C C C =
, (11.93)
In general, one can see that the response of each sample is a vector. However,
from Eq. (11.93), we can know Y is a tensor. To be consistent with the original form,
we can obtain the response of a test sample by using Y, denoted by
( )
( )
final n
Y mean Y = , or
( )
max(
n
Y ), or
( )
min( )
n
Y , (11.94)
where
( ) n
Y is the unfolding matrix of tensor Y in Eq. (11.93).
11.4 summAr y
In this chapter, we introduce tensor-based classifers and tensor canonical correla-
tion analysis and tensor partial least squares. First, the vector-based classifers are
extended to accept tensors as input, which is the multilinear extension of classifers.
is prohibited.
The results provide a unifed framework for classifer design. Among these clas-
sifers, we mainly introduce tensor Ho-Kashyap HK classifers, Tensor LS-SVM
and classifcation learning based on convex optimization programming. To obtain
the solutions of these learning algorithms, the alternating optimization is adopted.
Second, we also demonstrate tensor canonical correlation analysis and tensor partial
least squares. Tensor CCA and tensor PLS give a reformative framework of clas-
sical CCA and classical PLS. In this framework, these two tensor subspaces can
directly use the image matrix to extract the features instead of matrix-to-vector
transformation, which can effectively avoid the singularity of matrices occurring
in classical subspace analysis. Overall, these two tensor subspace analysis can ef-
fciently solve the small sample size problem. Although we describe more examples
of tensor analysis in this chapter, some tensor analysis such as tensor cluster analysis
is not involved because the ideas in this chapter are easily extended to tensor cluster
analysis and other classical statistical learning methods.
r eferences
Baek, J., & Kim, M. (2004). Face recognition using partial least squares components.
Bousquet, O., Boucheron, S., & Lugosi, G. (2004). Introduction to statistical learning
theory, Advanced Lectures on Machine Learning, LNCS 3176, 169-207.
Boyd, S., Kim, S. J., Vandenberghe, L., & Hassibi, A. (2006). A tutorial on geometric
programming. Optim Eng., 8(1), 67-127.
Boyd, S., & Vandenberghe, L. (2004). Convex optimization. Cambridge, UK:
Cambridge University Press.
Chen, S. C., Wang, Z., & Tian, Y. J. (2007). Matrix-pattern-oriented Ho-Kashyap
classifer with regularization learning. Pattern Recognition, 40(5), 1533-1543.
Duda, R. O., Hart, P. E., & Stock, D. G. (2001). Pattern classifcation (2nd ed.).
New York: Wiley.
Fukunaga, K. (1990). Introduction to statistical pattern recognition (2nd ed.). New
York, Academic Press.
Hardoon, D. R., Szedmak, S., & Taylor, J. S. (2004). Canonical correlation analysis:
An overview with application to learning methods. Neural computation, 16(12),
2639-2664.
is prohibited.
Helland, I. S. (1988). On structure of partial least squares regression. Communica-
tions in Statistics of Simulation and Computation, 17(2), 581-607.
Highleyman, W. H. (1962). Linear decision functions with application to pattern
recognition. Proceedings of the IRE, (Vol. 50, No. 1, pp. 1501-1514).
Hotelling, H. (1936). Relations between two sets of vitiates. Biometrika, 28, 321-
377.
Howland, P., & Park, H. (2004). Generalized discriminant analysis using the general-
ized singular value decomposition. IEEE Trans. on Pattern Analysis and Machine
Intelligence, 26(8), 995-1006.
Leski, J. (2003). HoKashyap classifer with generalization control. Pattern Rec-
ognition Lett., 24(14), 2281-2290.
Lewi, P. J. (1995). Pattern recognition, refections from a chemometric point of view.
Chemometrics and Intelligent Laboratory Systems, 28(1), 23-33.
Lobo, M. S., Vandenberghe, L., Boyd, S., & Lebret, H. (1998). Applications of
second-order cone programming. Linear Algebr. Appl., 284(1-3), 193-228.
Melzer, T., Reiter, M., & Bischof, H. (2003). Appearance models based on kernel
canonical correlation analysis. Pattern Recognition, 36(9), 1961-1971.
Muller, K.-R., Mika, S., Ratsch, G., Tsuda, K., & Scholkopf, B. (2001). An intro-
duction to kernel-based learning algorithms. IEEE Trans. Neural Networks, 12(2),
181-202.
Oskuldsson, A. H. (1988). PLS regression methods. Journal of Chemometrics,
2(3), 211-228.
Rannar, S., Lindgren, F., Geladi, P., & World, S. (1994). A PLS kernel algorithm
for data sets with many variables and fewer objects, Part I: Theory and algorithm.
Chemometrics and Intelligent laboratory System, 8(2), 111-125.
Sun, Q., Zeng, S., Liu, Y., Heng, P., & Xia, D. (2005). A new method of feature fusion
and its application in image recognition. Pattern Recognition, 38(12), 2437-2448.
Suykens, J. A. K., & Vandewalle, J. (1999). Least squares support vector machine
classifers. Neural Processing Letters, 9(3), 293-300.
Tan, Y., Shi, L., Tong, W., Hwang, G., & Wang, C. (2004). Multi-class tumour
classifcation by discriminant partial least squares using microarray gene expres-
sion data and assessment of classifcation models. Computational Biology and
Chemistry, 28(3), 235-244.
is prohibited.
Tao, D., Li, X., Hu, W., Maybank, S. J., & Wu, X. (2005). Supervised tensor learning.
Proceedings of the IEEE international conference on data mining (pp. 450-457).
Tao, D., Li, X., Wu, X., Hu, W., & Maybank, S. J. (2006). Supervised tensor learn-
ing. Knowledge and Information Systems, 13(1), 1-42.
Vandenberghe, L., & Boyd, S. (1996). Semidefnite programming. SIAM Rev.,
38(1), 49-95.
Vanderbei, R. (2001). Linear programming: foundations and extensions (2nd ed.).
Berlin: Springer.
Wang, Z., & Chen, S. C. (2007). New least squares support vector machines based
on matrix patterns. Neural Processing Letters, 26(1), 41-56.
Wold, H. (1982). Soft modelling: the basic design and some extensions. Systems
Under Indirect Observation: Causality and Prediction, Amsterdam, North Hol-
land, 1-54.
Wold, H. (1985). Partial least squares. In N. L. Johnson & S. Kotz (Eds.), Encyclo-
paedia of statistical sciences (pp. 581-591). New York: Wiley.
Ye, J., Janardan, R., Park, C. H., & Park, H. (2004). An optimization criterion
for generalized discriminant analysis on undersampled problems. IEEE Trans. on
Zheng, W., Zhou, X., Zou, C., & Zhao, L. (2006). Facial expression recognition
using kernel canonical correlation analysis (KCCA). IEEE Trans. Neural Networks,
17(1), 233-238.
Section III
Biometric Fusion
is prohibited.
Chapter XII
From Single Biometrics to
Multi-Biometrics
Abstr Act
In the past decades while biometrics attracts increasing attention of researchers,
people also have found that the biometric system using a single biometric trait
may not satisfy the demand of some real-world applications. Diversity of biometric
traits also means that they may have different performance such as accuracy and
reliability. Multi-biometric applications emerging in recent years are a big progress
of biometrics. They can overcome some shortcomings of the single biometric system
and can perform well in improving the system performance. In this chapter we
describe a number of defnitions on biometrics, categories and fusion strategies of
multi-biometrics as well as the performance evaluation on the biometric system. The
frst section of this chapter describes some concepts, motivation and justifcation of
multi-biometrics. Section 12.2 provides some defnitions and notations of biometric
and multi-biometric technologies. Section 12.3 is mainly related to performance
evaluation of various types of biometric systems. Section 12.4 briefy presents
research and development of multi-biometrics.
12. 1 Introduct Ion
As mentioned in previous chapters, biometric technologies play an important role
in access control and other systems that depend on secure personal authentication.
From Single Biometrics to Multi-Biometrics 255
is prohibited.
The fact that biometrics may possess excellent properties such as universality
(every person has biometric traits), uniqueness (generally, no two people have
identical biometric traits), permanence (most biometric traits do not vary over time),
collectability (biometric traits can be measured quantitatively) and good performance
(biometric technologies can achieve accurate results under varied environmental
circumstances) (Ross & Jain, 2004; Jain, Ross, & Prabhakar, 2004) provides a solid
base for these systems. Indeed, biometric technology is a methodology to achieve
fast, user-friendly authentication with high accuracy. As mentioned in Chapter I,
compared with biometric systems, traditional security systems, such as passwords
or tokens-based methods, have some serious disadvantages.
Biometric technologies have many applications (Jain, Bolle, & Pankanti, 1999;
Zhang & Jain (Eds.), 2006; Wayman, 2001; Bolle, Connell, Pankanti, Ratha, &
Senior, 2004; Herzog & Reithiger, 2006; Jain & Ross, 2004; Jain, 2003). Biomet-
rics can be incorporated in solutions to provide for Homeland Security including
applications for improving airport security, strengthening border management
control, in visas and in preventing ID theft. Biometrics can be also applied to
secure electronic banking, investing and other fnancial transactions, enterprise-
wide network security infrastructures, retail sales, law enforcement, and health and
social services. Biometrics can also be integrated with other technologies such as
encryption keys or smart cards to produce a hybrid security system. This way of
exploiting biometrics is also called two-factor authentication, please refer to web
site (http://www.answers.com/topic/two-factor-authentication). As shown in Chapter
I, biometric applications can be categorized into several categories.
Varieties of biometric traits can be individually applied for personal authentication
(Zhang, Jing, & Yang, 2005; Zhang & Jain (Eds.), 2004; Zhang, 2004); however, the
biometric system using a single biometric trait usually suffers from some problems
such as unsatisfactory accuracy, spoof attacks, and restricted degrees of freedom
(Bubeck, 2003). For example, manual workers with damaged or dirty hands may
not be able to provide high-quality fngerprint images. In this case, fngerprint au-
thentication seems not to be a good means for authenticating personal identity. For
an iris identifcation system, the existing registration failure risk would reduce the
reliability of the system. For a biometric system using speech, some factors, such
as ambient noise, changes in behavioral attributes of the voice, and voice change
due to aging, will affect the systems performance. For a biometric system using
face images, some challenges, such as variations in facial expression, pose and
lighting, will limit the systems performance. Ones keystroke trait and signature
trait may vary to some extent and also bring side effects into the single biometric
system using keystrokes or signature traits. All these examples imply that the single
biometric system may not be guaranteed to provide a high accuracy.
is prohibited.
In the past some means to improve biometric systems reliability have been
investigated. It has been proven that the simultaneous use of two or more biometric
traits of the same identity is an effective way to enhance the performance of the
biometric system (Jain, Prabhakar, & Chen, 1999; Ross & Jain, 2003; Prabhakar &
Jain, 2002; Zhang & Jain (Eds.), 2006). Biometric systems that integrate multiple
biometric traits are called multi-biometric systems or multimodal biometric system.
Biometric systems that use only one biometric trait are called single biometric
systems. It is well recognized that the combination of typical physiological or be-
havioral traits make biometric systems more accurate. Moreover, it is reported that
non-conventional traits, i.e. the so-called soft traits such as color of skin, color of
iris and height of the subject might also be integrated with conventional biometric
traits to enhance the overall performance of the system. Different from soft traits,
the term hard traits is used to denote conventional biometric traits such as iris,
face and palm print. By far, many multi-biometric techniques have been proposed
and evaluated (Jain, Bolle, & Pankanti, 1999; Ross & Jain, 2003; Jain, Pankanti, &
Chen, 1999; Hong & Jain, 1998; Kresimir & Mislav, 2004; Snelick, Uludag, & Mink,
2005). The promising results of multi-biometric systems encourage engineers and
practitioners to devote themselves to the development of biometric systems. The
signifcance of multi-biometrics is also partially refected by the following statement:
Multi modal technology makes biometric work (Aurora Defense, 2002).
One of the methodological rationales of multi-biometrics is as follows: since
different biometric traits such as the iris and the fngerprint can be considered
independent of each other, the combination of them means that more information
is used, which is theoretically helpful to improve the performance of biometric
systems. Indeed, higher accuracy, which may be considered one of the essential
goals of biometric technology, is also the primal reason of the appeal of multi-
biometric systems.
The multi-biometric system will also provide strong feasibility and reliability,
which can be explained in twofold: First, multi-biometric technology is able to
improve the anti-spoofng performance of the system because it is much more dif-
fcult for an intruder to simultaneously spoof multiple biometric traits. Second, in
the case where one biometric trait of an individual is not accessible (due to illness
or injury ), a multi-biometric system may still work, as long as the other traits are
accessible. However, a single biometric system cannot work in the case where the
user does not possess the required particular biometric trait. Varieties of studies
and experiments have shown that multi-biometrics does perform better than single
biometrics. For example, biometrics using face and fngerprint (Snelick et al., 2005;
Hong & Jain, 1998), iris and face (Wang, Tan, & Jain, 2003), ear and face (Chang,
Bowyer & Sarkar, 2003), palm print and hand geometry (Kumar, Wong, Shen, &
is prohibited.
Jain, 2003), or face and speech data (Ben-Yacoub, Abdeljaoued, & Mayoraz, 1999)
all are capable to obtain a higher accuracy than the corresponding single biometric
technologies.
12.2 bI ometr Ic And mul t I-bI ometr Ic f us Ion:
def InIt Ion And not At Ion
12.2.1 Defnitions
Total Risk
Let R be the set that consists of similarity values representing genuine users. The
total risk of a biometric system using a single trait can be defned as E(R) = C
FAR
*
FAR(R) + C
FRR
* FRR(R), where C
FAR
and C
FRR
are respectively the costs of falsely
accepting an imposter and of falsely rejecting an genuine user. It is clear that if C
FAR

= C
FRR
, then the total risk E(R) is the same as the total error, FAR(R) = FRR(R). We
can defne the minimum total risk as follows (Hong, Jain, & Pankanti, 2000):
min
( ) min ( ( ) ( ))
t FAR FRR
E R C FAR R C FRR R = +
, (12.1)
where t denotes the threshold. It has been theoretically demonstrated that a multi-
biometric system which fuses multiple biometric traits at the matching score level
or decision level can obtain lower total risk than the corresponding single biometric
system using a single biometric trait (Hong, Jain, & Pankanti, 2000).
Verifcation
A verifcation process can be described formally as follows: given an input feature
vector X
G

of the single biometrics and the claimed identity of the user, determine if
X
G
belongs to class
1
or
2
, where
1
means that the claim is true (a genuine user)
and
2
means that the claim is false (an impostor). Generally, we should match
X
G
against the registered biometric template of the claimed identity, to determine
whether the user is a genuine user. If the similarity between feature vector X
G
and
the registered biometric template of the claimed identity is greater than a predefned
threshold, we can consider that the user is a genuine user; otherwise the user is
considered to be an impostor.
is prohibited.
Identifcation
Identifcation can be described as follows: given an input feature vector X
G
of the
single biometric trait of a user, determine the identity of the user. Generally, we
can compute the similarities between X
G
and the registered biometric templates of
different identities and select the identity having the maximum similarity as the
identity of the user.
FAR, FRR
False reject rate (FRR) of a genuine user n is defned as
Number of rejected verification attempts for a genuine user n
( )
Number of all verification attempts for the same genuine user n
FRR n =
.
The overall FRR for N genuine users is defned as follows:
1
1
( )
N
n
FRR FRR n
N
=
=

.
False accept rate (FAR) of an imposter n is defned as
Number of accepted verification attempts for an imposter
( )
Number of all verification attempts for the same imposter
n
FAR n
n
=
.
The overall FAR for N imposters is defned as follows:
1
1
( )
N
n
FAR FAR n
N
=
=

.
FMR
In the case of identifcation, False Match Rate (FMR) is defned as the rate that
the biometric system incorrectly declares a successful match between the input
pattern and a nonmatching pattern in the database. In the case of verifcation,
FMR refers to the rate of the impostor whose (similarity) score is greater than the
decision threshold.
FNMR
In the case of identifcation, False Non-match Rate (FNMR) is defned as the rate
that the biometric system incorrectly declares failure of match between the input
is prohibited.
pattern and a matching pattern in the database. In the case of verifcation, FNMR
refers to the rate of the genuine user whose (similarity) score is smaller than the
decision threshold.
FTAR
FTAR refers to failure to acquire rate (FTAR). FTARs of practical biometric systems
are affected by the nature of the capture device and the biometric traits as well as
the cooperation of the user.
FTER
FTER refers to failure to enroll rate (FTER) in the system register (enroll) stage.
Usually failure to enroll happens when the obtained enroll data are considered
invalid or of poor quality.
Receiver Operating Characteristic (ROC)
ROC represents a graphic means for assessing the ability to discriminate between
genuine users and imposters. The ROC plot can be obtained by graphing the values
of FAR and FRR, changing the threshold implicitly. Note that both the theoretical
and the experimental studies show that in a biometric system in a practical situ-
ation attempts for achieving a low FAR by adjusting the predefned threshold for
verifcation usually should at the cost of a high FRR (Golfarelli, Maio, & Maltoni,
1997). In other words, it is almost impossible for a certain system to simultaneously
reduce the FAR and FRR of the system only by adjusting the predefned threshold
for verifcation. For real-world applications, it is an important issue to select an ap-
propriate threshold and to obtain the best trade-off between FAR and FRR.
EER
Equal error rate (EER) refers to the rate at which both accept and reject errors are
equal. The EER is obtained from the ROC plot by taking the point where FAR and
FRR have the same value. The ERR is usually used to compare performances of two
systems. The lower the EER is, the more accurate the system is considered to be.
12.2.2 Categories of Multi-Biometrics
Though the formal defnition and categories of types of biometric fusion systems is
still an active topic, INCITS (InterNational Committee for Information Technology
is prohibited.
Standards) Secretariat, Information Technology Industry Council (ITI) classifes
them into three primary categories, defned as follows:
A. Multi-Modal. Multi-modal biometrics means the use of two or more funda-
mentally different types of raw biometric samples, such as face image and
fngerprint image data. This includes cases where the same body part is
sampled in two fundamentally different ways. Examples include a hand that
can be sampled as a palm print or an outline (hand geometry).
B. Multi-Instance. Multi-instance biometrics means the use of the same type of
raw biometric sample and processing on multiple instances of similar body
parts, such as two fngers, or two irises.
C. Multi-Algorithmic. Multi-algorithmic biometrics means the use of multiple
algorithms of processing on the same raw biometric samples, such as minutia
and ridge pattern algorithms applied to the same fnger image data.
Raw sample means the data captured by a biometric capture system before any
processing by biometric algorithms that perform either registration (determination
of landmarks or features) or template creation.
Multi-biometrics can also be categorized in other ways. For example, as presented
in Section 12.2.3, multi-biometrics can also be categorized depending on the stage
in which of the authentication process the information of different biometric traits
is combined. In addition, two terms bimodal-biometrics and multi-modal biometrics
are also used to categorize biometrics in some cases where people care how many
biometric traits are used in the multi-biometric system. While bimodal-biometrics is
usually referred to as the biometric system that uses simultaneously two biometric
traits, multi-modal biometrics can be referred to as the biometric system that uses
simultaneously not less than three biometric traits. Indeed, bimodal biometrics can
be regarded as the simplest case of generalized multi-modal biometrics.
A single biometric system can be presented by the fowchart as shown in Figure
12.1, while a multi-biometric system can be presented by the fowchart as shown
in Figure 12.2. The fusion process as shown in Figure 12.2 implements the
information fusion of multi-modal biometrics. The meanings of the terms in the
fowchart are explained as follows:
The StoredBioRecordis the template or image data enrolled for each user.
The DemographicRecord is the user specifc information deemed advanta-
geous for decision making in the case where the biometric trait of the subject is
not accessible.
The DecisionRecordis the information created and transmitted after the deci-
sion process.
is prohibited.
Figure 12.1. Flowchart of a single biometric system
Figure 12.2. Flowchart of a multi-modal biometric system
Biometric
Process
Decision
Record
Capture Demographic
Record
Captured
Sample
Stored Bio
Record
Demographic
Record
Biometric
Process A
Capture A
Record A
Captured
Sample A
Stored Bio
Record A
Demographic
Record
Stored Bio
Record B
Captured
Sample B
Biometric
Process B
Record B
Capture B
A
Fusion
Process
B
Decision
Record
12.2.3 Fuse information of Multi-Biometrics at Different Levels
If the systems are categorized according to the stage in which of the authentication
process the information from the different sensors is combined, there are three
categories of fusion strategies: fusion at the feature level, fusion at the matching
is prohibited.
score level, and fusion at the decision level. Fusion at the feature level means im-
mediate data integration at the beginning of the authentication process whereas
fusion at the decision level represents data integration at the end of the process.
Fusion at the feature level can be described as follows. The data obtained from
each sensor is used to compute a feature vector. As the feature extracted from one
biometric trait is independent of that extracted from the other, it is reasonable to
concatenate the two vectors into a single new vector for performing multi-biometrics
based personal authentication. Note that the new feature vector now has a higher
dimensionality than the original feature vector generated from each sensor. Feature
reduction techniques may be employed to extract useful features from the new
feature vector. This class of fusion is illustrated in Figure 12.3.
Fusion at the matching score level can be depicted as follows. Each subsystem
using one biometric trait of the multi-biometric system provides a matching score
indicating the proximity of the feature vector with the template vector. These scores
can be combined to assert the veracity of the claimed identity. Fusion at the match-
ing score level is illustrated in Figure 12.4.
Fusion at the decision level for verifcation can be implemented using the
following procedures. First, each sensor captures one of multiple biometric traits
Figure 12.3. Fusion at the feature level
Figure 12.4. Fusion at the matching score level
Feature
Matching
Biometric
traits
Biometric
traits
Feature
Extraction
Feature
Extraction
Decision
Making
Score
fusion
Accept
or
reject
Feature
Matching
Biometric
traits
Biometric
traits
Feature
Extraction
Feature
Extraction
Feature
Matching
Decision
Making
Accept
or
reject
Feature
fusion
is prohibited.
Figure 12.5. Fusion at the decision level
Biometric
traits
Biometric
traits
Feature
Extraction
Feature
Extraction
Decision
Making
Feature
Matching
Feature
Matching
Decision
fusion
Accept
or
reject
Decision
Making
and the resulting feature vectors are individually classifed into the two decisions
accept or reject the claimed identity. Then a scheme that exploits the known
decisions to make the fnal decision is used. Fusion at the decision level is illustrated
in Figure 12.5.
In the feld of multi-biometrics, a number of studies of feature level fusion
(Chang, Bowyer, Sarkar, & Victor, 2003; Ross & Govindarajan, 2005; Kong, Zhang,
& Kamel, 2006; Kober, Harz, & Schiffers, 1997; Gunes & Piccardi, 2005; Jing,
Yao, Zhang, Yang, & Li, 2007), matching score level fusion (Dass, Nandakumar,
& Jain, 2005; Schmid, Ketkar, Singh, & Cukic, 2006) and decision level fusion
(Verlinde & Cholet, 1999; Chatzis, Bors, & Pitas, 1999; Osadciw, Varshney, &
Veeramachaneni, 2003; Gkberk & Akarun, 2006) have been made. Although
fusion of multi-biometrics are generally recognized as three classes as described
above, in real-world applications of multi-modal biometrics it is possible that the
fusion process is simultaneously involved in different levels such as in both the
matching score level and the decision level (Jain, Chen, & Demirkus, 2007).
12.3 Perform Ance evAlu At Ion of bI ometr Ic
t echn Iques
While various types of biometric techniques have been developed, people fnd that
biometric traits have different natures and these techniques also have different
performance. As shown in Table 12.1, biometric techniques can be scored in terms
of the nature of the biometric trait such as universality (1), distinctiveness (2),
permanence (3), collectability (4), performance (5), and acceptability (6).
From Table 12.1, we can see that almost no one biometric trait simultaneously
have high distinctiveness, permanence, collectability, performance and
acceptability. Table 12.2 shows accuracies of state of art of biometric recognition
is prohibited.
Table 12.1. Performance scores of different biometric techniques
H, M and L represent high, medium and low, respectively
universality distinctiveness permanence collectability performance acceptability
Face H L M H L H
Hand geom-
etry
M M M H M M
Fingerprint M H H M H M
Iris H H H M H L
Retina H H M L H L
Signature L L L H L H
Voice M L L M L H
Table 12.2. State of art of biometric recognition systems
EER FAR FRR
Number of
Subjects
Comment Reference
Face n.a. 1% 10% 37437
Varied lighting, indoor/out-
door
FRVT
Fingerprint n.a. 1% 0.1% 25000
US Government operational
data
FpVTE
Fingerprint 2% 2% 2% 100
Rotation and exaggerated
skin distortion
FVC
Hand
geometry
1% 2% 0.1% 129
With rings and improper
placement
(2005)
Iris < 1% 0.94% 0.99% 1224 Indoor environment ITIRT
Iris 0.01% 0.0001% 0.2% 132 Best conditions NIST
Voice 6% 2% 10% 310
Text independent, multilin-
gual
NIST
systems. We should point out aside the nature of the biometric trait, the accuracy of
the biometric system is also directly related to the sensor, scalability of the user, and
the algorithm. As a result, Table 12.2 does not mean that when we use a biometric
trait shown in this table we can obtain the same accuracy. Indeed, we can consider
that the accuracy shown in Table 12.2 was obtained under the condition of zero
FTAR and FTER. It is probably that real-world application of the biometric trait
usually produces a lower accuracy than that shown in Table 12.2 because of nonzero
is prohibited.
FTAR and FTER and other uncertain factors such as lack of accessibility to the
biometric trait due to illness or injury. Jain, Pankanti, Prabhakar, Hong, and Ross
(2004) indicated that a large gap exists between the current biometric technology
and the performance requirement.
12.4 r ese Arch And develo Pment of
mul t I-bI ometr Ics
12.4.1 Basis and Potential of Multi-Biometrics
The basis and potential of multi-biometrics can be shown from theory or practical
perspective. We frst briefy present some signifcant theory research on the basis
and potential of multi-biometrics as follows. Kittler, Hatef, Duin, and Matas (1998)
studied the problem of combining classifers that used different representations
of the patterns to be classifed and developed a common theoretical framework
for classifer combination. They also showed that under different assumptions
and using different approximations they could derive the commonly used classi-
fer combination schemes such as the product rule, sum rule, min rule, max rule,
median rule, and majority voting. Golfarelli Maio and Maltoni (1997) studied the
problem of performance evaluation in biometric verifcation systems. They derived
two statistical expressions for theoretically calculating FRR and FAR. They also
demonstrated the fundamental relation between FRR and FAR. Hong, Jain and
Pankanti (2000) theoretically proved that it was possible to improve performance of
the biometric system by integrating multiple biometric traits. Kuncheva, Whitaker,
Shipp, and Duin (2000) studied the limits on the major voting accuracy produced
by the combination of dependent classifers. Jain, Nandakumar, and Ross (2005)
analyzed and compared the performance of different normalization techniques and
fusion rules. Dass, Nandakumar and Jain (2005) proposed an optimal framework
for combining the matching scores from multiple modalities using the likelihood
ratio statistic computed using the generalized densities estimated from the genu-
ine and impostor matching scores. A theoretical analysis of a novel approach was
proposed based on the dynamic selection of matching scores (Tronci, Giacinto,
& Fabio, 2007). Such a selector aims at choosing, for each user to be authenticated,
just one of the scores produced by the different biometric systems available. Dass,
Zhu,and Jain (2006) analyzed and discussed the problem of constructing conf-
dence regions based on the ROC curve for validating the claimed performance
levels and the problem of determining the required number of biometric samples
needed to establish confdence regions of pre-specifed width for the ROC curve.
Osadciw, Varshney and Veeramachaneni (2003) explored optimum fusion rules for
is prohibited.
multimodal biometrics. Sim, Zhang, Janakiraman, and Kumar (2007) presented the
theory, architecture and implementation of the multimodal biometric verifcation
system that continuously verifes the presence of a logged-in user. They also argued
that the usual performance metrics of false accept and false reject rates were insuf-
fcient yardsticks for continuous verifcation and proposed new metrics. Oermann,
Scheidat, and Vielhauer (2006) presented a theoretical concept of a methodology
to improve those fusions and strategies independently of their application levels.
They showed that the fusion could be potentially improved by extracting and
merging certain semantic information and integrating it as additional knowledge
(e.g. metadata) into the fusion process. Zhang and Zuo (2007) and Xiao (2007)
summarized intelligent computation technologies in the felds of biometrics from
different view of points.
There is also a variety of practical research on multi-biometrics. Chang, Bow-
yer, Sarkar, and Victor (2003) combined the ear and face image to perform identity
recognition and could obtain the accuracy of 90.9%, whereas the accuracies of the
ear-based and face-based identity recognition are 71.6% and 70.5%, respectively.
Other examples for illustrating that the use of multi-modal biometrics brings
performance improvement include the combination of, fngerprint and face (Hong
& Jain, 1998), iris and face (Wang, Tan, & Jain, 2003), speech and face (Ben-Ya-
coub et al., 1999; Teoh, Samad, & Hussain, 2002), visual and acoustic signals for
command-word recognition (Kober, Harz , & Schiffers, 1997). etc. In addition,
studies on the multi-instance biometrics such as personal authentication using two
irises (Wu, Wang, Zhang, & Qi, 2007), two palm prints (Wu, Wang, & Zhang,
2007), hand geometry and texture (Kumar & Zhang, 2006; Kumar, Wong, Shen, &
Jain, 2003), the combination of shape contexts and local features of signature (Li,
Zhang, & Wang, 2006) also suggested clear performance improvement. Jain, Chen
and Demirkus (2007) recently made a signifcant progress in fngerprint verifcation.
They proposed to verify fngerprint using level 3 features (ridge fow and pattern
type, minutia points and pores as well as ridge contours) and achieved a relative
performance gain of 20 percent in terms of EER over the Level 2 matcher.
Besides the integration of different types of biometric technologies and the
combination of multi-instances can improve the performance of the biometric
system; the advancement in the sensor and algorithm design also allows the biometric
system to produce a higher accuracy. Take face recognition as an example, while
the 2002 Face Recognition Vendor Test (FRVT 2002) reported that the best 2002
face recognition probability of verifcation using a single face image with controlled
illumination was 90 percent (Philips, Grother, Micheals, Blackburn, Tabassi, &
Bone, 2002), 2006 Face Recognition Vendor Test (FRVT 2006) obtained a FRR of
0.01 at a FAR of 0.001 by using very high-resolution still images and Viisage (V-3D-
is prohibited.
n algorithm) on the 3D images. Development of palm print verifcation is another
example. Zhang, Kong, You and Wong (2003) developed a CCD-based plamprint
device and collected a palm print image database, 7,752 images from 386 different
palms. Using this database, they obtained a 98 percent genuine acceptance rate
and a 0.04 percent false acceptance rate with the threshold 0.3425. Wu, Zhang and
Wang (2006) extracted the palm line feature, one of the most important feature of
the palm print image. Palm print verifcation using the obtained palm lines obtained
the EER of 0.0084. In addition fusion of phase and orientation information of the
palm print image also can obtain a satisfactory accuracy (Wu, Zhang, & Wang,
2006; Kong, Zhang, & Kamel, 2006).
12.4.2 Use and Application of Multi-Biometrics

In recent years many governments have increasing concerns about the multi-biometric
technology. In May 2005 the German Upper House of Parliament approved the
implementation of the ePass, which enables the passport issued to all German citizens
to contain biometric traits (please refer to: http://www.itsmig.de/best_practices/eP-
ass_en.php). The issued ePass contains a chip that holds a digital photograph and
one fngerprint from each hand, usually of the index fngers, though others may be
used if these fngers are missing or have extremely distorted prints. In Germany, the
new work visas will also include fngerprinting, iris scanning, and digital photos
(please refer to: http://www.flipinoexpress.com/19/19_news.html). In Israel the
biometric technology has been used for several years. For example, at the border
crossing points from Israel to the Gaza Strip and West Bank, authorized Palestin-
ians can pass through the turnstiles through the authorization check based on their
biometric traits and the held smartcard with stored information of fngerprints,
facial geometry and hand geometry. At peak periods more than 15,000 people an
hour pass through the gates. In Israel another example of biometric application is
in the Tel Aviv Ben Gurion Airport. This airport has a frequent fyer's fast check-in
system based on the use of a smartcard that holds information of the hand geometry
and fngerprints of the holders. In Iraq, biometric technologies are also being used
extensively (please refer to: http://www.csc.com/solutions/security/casestudies/2829.
shtml). For example, during account creation, the collected biometric information can
be inputted into a central database. Also, additional information such as individual
personal history can be added to each account record. Besides the governments
mentioned above, the United States government has also become a strong advocate
of biometrics in recent years.
is prohibited.
r eferences
Aurora Defense LLC. (2002). PRWeb Press Release. From http://www.prweb.com/
releases/2002/2/prweb33800.php.
Ben-Yacoub, S., Abdeljaoued, Y., & Mayoraz, E. (1999). Fusion of face and speech
data for identity verifcation. IEEE Trans. Neural Networks, 10(5), 1065-1075.
Bubeck, U. M. (2003). Multibiometric authentication- an overview of recent
developments. San Diego: San Diego State University.
Chang, K., Bowyer, K. W., Sarkar, S., & Victor, B. (2003). Comparison and combi-
nation of ear and face images in appearance-based biometrics. IEEE Trans. Pattern
Chatzis, V., Bors, A. G., & Pitas, I. (1999). Multimodal decision-level fusion for
person authentication. IEEE Trans. on SMC-A, 29 (6), 674-680.
Dass, S. C., Nandakumar, K., & Jain, A. K. (2005). A principled approach to score
level fusion in multimodal biometric systems. Proc. of Audio- and Video-based
Biometric Person Authentication (AVBPA) (pp. 1049-1058).
Dass, S., Zhu, Y., & Jain, A. (2006). Validating a biometric authentication system:
sample size requirements. IEEE Trans. Pattern Anal. Machine Intell., 28(12), 1902-
1913.
Golfarelli, M., Maio, D., & Maltoni, D. (1997). On the error-reject tradeoff in
biometric verifcation systems. IEEE Trans. Pattern Anal. Machine Intell., 19(7),
786-796.
Gkberk, B., & Akarun, L. (2006). Comparative analysis of decision-level fusion
algorithms for 3D face recognition. Proc. of International Conference on Pattern
Gunes, H., & Piccardi, M. (2005). Affect recognition from face and body: early
fusion vs. late fusion. Proc. of 2005 IEEE International Conference on , Systems,
Man and Cybernetics (pp. 3437-3443).
Herzog, G., & Reithiger, N. (2006). The SmartKom: foundations of multimodal
dialogue systems (pp. 43-58). New York: Springer-Verlag New York Inc.
Hong, L., & Jain, A. (1998). Integrating faces and fngerprints for personal identi-
fcation. IEEE Trans. Pattern Anal. Machine Intell., 20(12), 1295-1307.
Hong, L., Jain, A., & Pankanti, S. (2000). Can multibiometrics improve perfor-
mance?. Proceedings AutoID (pp. 59-64).
is prohibited.
Jain, A. (2003, November). Multimodal user interfaces: Whos the user? Invited
talk in Fifth International Conference on Multimodal Interfaces. Vancouver, Brit-
ish Columbia, Canada.
Jain, A., Bolle, R., & Pankanti, S. (1999). Biometrics: personal identifcation in
networked society. Boston Hardbound: Kluwer Academic Publishers.
Jain, A. K., Chen, Y., & Demirkus, M. (2007). Pores and ridges: high-resolution
fngerprint matching using level 3 features. IEEE Trans. Pattern Anal. Machine
Intell., 29(1), 15-27.
Jain, A. K., Jain, K., & Ross, A. (2005). Score normalization in multimodal biometric
systems. Pattern Recognition, 38, 2270 2285.
Jain, A. K., Prabhakar, S., & Chen, S. (1999). Combining multiple matchers for a
high security fngerprint verifcation system. Pattern Recognition Letters, 20(11-
13), 1371-1379.
Jain, A. K., Pankanti, S., Prabhakar, S., Hong., L., & Ross, A. (2004). Biometrics:
a grand challenge. Proc. of the 17th International Conference on Pattern Recogni-
tion, 2, 935942.
Jain, A., & Ross, A. (2004). Multibiometric systems. Communications of the ACM,
Special Issue on Multimodal Interfaces, 47(1), 34-40.
Jain, A. K., Ross, A., & Prabhakar, S. (2004). An introduction to biometric recognition.
IEEE Trans. on Circuits and Systems for Video Technology, Special Issue on Im-
age- and Video-Based Biometrics, 14(1), 4-20.
Jing, X. Y., Jing, Y. F., Zhang, D., Yang, J. Y., & Li, M. (2007). Face and palmprint
pixel level fusion and Kernel DCV-RBF classifer for small sample biometric rec-
ognition. Pattern Recognition, 40(11), 3209-3224.
Kittler, J. Hatef, M., Duin, R. P. W., & Matas, J. (1998). On combining classifers.
Kober R., Harz U., & Schiffers J. (1997). Fusion of visual and acoustic signals for
command-word recognition, Proc. of International Conference on Acoustics, Speech
and Signal Processing (pp. 1495-1497).
level fusion. Pattern Recognition, 39(3), 478-487.
Kresimir, D., & Mislav, G. (2004). A survey of biometric recognition methods.
Proceedings of the 46th International Symposium Electronics in Marine(ELMAR-
2004) (pp.16-18).
is prohibited.
Kumar, A., Wong, D. C. M., Shen, H. C., & Jain, A. K. (2003). Personal verifcation
using palmprint and hand geometry biometric. Proceedings of the fourth Interna-
tional Conference on Audio-and Video-Based Biometric Person Authentication
(AVBPA) (pp. 668-678).
Kumar, A., & Zhang, D. (2006). Personal recognition using hand shape and texture.
IEEE Trans. Image Processing, 15(8), 2454-2461, 2006.
Kuncheva, L. I., Whitaker, C. J., Shipp, C. A., & Duin, R. P. W. (2000). Is indepen-
dence good for combining classifers? Proc. International Conference on Pattern
Li, B., Zhang, D., & Wang, K. (2006). Online signature verifcation by combining
shape contexts and local features. Pattern Recognition, 39(11), 2244-2247.
Oermann, A., Scheidat, T., Vielhauer, C., & Dittmann, J. (2006). Semantic fusion
for biometric user authentication as multimodal signal processing. Proceedings
of international workshop Multimedia content representation, classifcation and
security (MRCS) (pp. 546-553).
Osadciw, L. A., Varshney, P. K., & Veeramachaneni, K. (2003). Multisensor
surveillance systems: a fusion perspective. Boston: Kluwer Academic Publishers.
Prabhakar, S., & Jain, A. K. (2002). Decision-level fusion in fngerprint verifcation.
biometrics. Proc. of SPIE Conference on Biometric Technology for Human Iden-
tifcation II (pp. 196-204).
Ross, A., & Jain, A. (2003). Information fusion in biometrics. Pattern Recognition
Letters, 24(13), 2115-2125.
Ross, A., & Jain, A. K. (2004). Multimodal biometrics: An overview. Proc. of the
12th European Signal Processing Conference (pp. 1221-1224).
Schmid, N. A., Ketkar, M., Singh, H., & Cukic, B. (2006). Performance analysis
of iris-based identifcation system at the matching score level. IEEE Trans. on
Information Forensics and Security, 1(2), 154-168.
Sim, T., Zhang, S., Janakiraman, R., & Kumar, S. (2007). Continuous verifcation
using multimodal biometrics. IEEE Trans. Pattern Analysis and Machine Intel-
ligence, 29(4), 687-700.
is prohibited.
Snelick, R., Uludag, U., Mink, A., Indovina, M., & Jain, A. (2005). Large-scale
evaluation of multimodal biometric authentication using state-of-the-art systems.
Teoh, A. B. J., Samad, S. A., & Hussain, A. (2002). Decision fusion comparison
for a biometric verifcation system using face and speech. Malaysian Journal of
Computer Science, 15(2), 17-27.
Tronci, R., Giacinto, G., & Roli, F.(2007). Dynamic score selection for fusion of
multiple biometric matchers. 14th International Conference on Image Analysis and
Processing (ICIAP 2007) (pp. 15-22).
Verlinde, P., & Cholet, G. (1999). Comparing decision fusion paradigms using k-NN
based classifers, decision trees and logistic regression in a multi-modal identity
verifcation application. Proceedings of International Conference on Audio and
Video-Based Biometric Person Authentication (AVBPA) (pp. 188-193).
Wang, Y., Tan, T., & Jain, A. K. (2003). Combining face and iris biometrics for
identity verifcation. Proc.of International Conference on Audio and Video-Based
Wayman, J. L. (2001). Fundamentals of biometric authentication technologies.
International Journal of Image and Graphics, 1(1), 93-113.
Wu, X., Wang, K., & Zhang, D. (2007). Automated personal authentication using
both palmprints. Proceedings of the International Conference on Entertainment
Computing (pp. 450-453).
Wu, X., Wang, K., Zhang, D., & Qi, N. (2007). Combining left and right irises for
personal authentication. Proceedings of the 6th International Conference Energy
Minimization Methods in Computer Vision and Pattern Recognition (pp. 145-
152).
Wu, X., Zhang, D., & Wang, K. (2006). Fusion of phase and orientation information
for palmprint authentication. Pattern Anal. Appl., 9(2-3), 103-111.
Wu, X., Zhang, D., & Wang, K. (2006). Palm-line extraction and matching for
personal authentication. IEEE Trans. SMC-A, 36(5), 978-987.
Xiao, Qinghan (2007). Biometricstechnology, application, challenge, and
computational intelligence solutions. IEEE Computational Intelligence Magazine,
2(2), 5-25.
Zhang, D. (2004). Palmprint Authentication. Boston: Kluwer Academic Publish-
ers.
is prohibited.
Zhang, D., & Jain, A. K. (Eds.) (2004). Biometric authentication. New York:
Springer-Verlag New York, Inc.
Zhang, D., & Jain, A. (Eds.) (2006). Advances in biometrics. New York: Springer-
Verlag New York, Inc.
Zhang, D., Jing, X., & Yang, J. (2005). Biometric Images Discrimination (BID)
Technologies. Hershey: Idea Group Publishing.
Zhang, D., Kong, W., You, J., & Wong, M. (2003). Online palmprint identifcation.
Zhang, D., & Zuo, W. (2007). Computational intelligence-based biometric technolo-
gies. IEEE Computational Intelligence Magzine, 26-36.
Feature Level Fusion 273
is prohibited.
Chapter XIII
Feature Level Fusion
Abstr Act
This chapter introduces the basis of feature level fusion and presents two feature level
fusion examples. As the beginning, Section 13.1 provides an introduction to feature
level fusion. Section 13.2 describes two classes of feature level fusion schemes. Sec-
tion 13.3 gives a feature level fusion example that fuses face and palm print. Section
13.4 presents a feature level fusion example that fuses multiple feature presentations
of a single palm print trait. Finally, Section 13.5 offers brief comments.
13.1 Introduct Ion
Fusion at the feature level means that the combination of different biometric traits
occurs at an early stage of the multi-biometric system (Chang, Bowyer, & Sarkar,
2003; Gunatilaka & Baertlein, 2001; Gunes & Piccardi, 2005; Jain, Ross, & Prab-
hakar, 2004; Kober, Harz, & Schiffers, 1997; Kong, Zhang, & Kamel, 2006; Ross &
Govindarajan, 2005; Ross & Jain, 2003). The traits fused at the feature level will be
used in the matching and decision-making modules of the multi-biometric system
to obtain authentication results. As pointed out by Jain, Ross and Prabhakar (2004)
and other researchers (Choi, Choi, & Kim, 2005; Ratha, Connell, & Bolle, 1998;
Singh, Vatsa, Ross, & Noore, 2005), it is likely that the integration at the feature
level is able to produce a higher accuracy than fusion at the matching score level
and fusion at the decision level. This is because the feature representation conveys
richer information than the matching score or the verifcation decision (i.e. accept
is prohibited.
or reject) of a biometric trait. In contrast, the decision level is so coarse that much
information loss will be caused (Sim, Zhang, Janakiraman, & Kumar, 2007). Fol-
lowing are some examples of fusion at the feature level in the feld of biometrics.
Chang, Bowyer and Sarkar (2003) fused appearance traits of the face and ear at
the feature level. They concatenated the traits of the face and ear and exploited the
resultant new one-dimensional data to perform personal authentication. To reduce
the dimensionality of the new data, they extracted features from the new data. Ross
and Govindarajan (2005) discussed fusion at the feature level in the following dif-
ferent scenarios such as fusion of PCA and LDA coeffcients of face images and
fusion of face and hand modalities. Kong, Zhang and Kamel (2006) fused the phase
information of different Gabor fltering results at the feature level to perform palm
print identifcation. Gunatilaka and Baertlein (2001) proposed a feature-level fusion
approach for fusing data generated from non-coincidently sampled sensors. Other
feature level fusion examples include fusion of visual and acoustic signals (Kober,
Harz, & Schiffers, 1997), the fusion of face and body information (Gunes & Pic-
cardi, 2005), fusion of face and fngerprint (Rattani, Kisku, Bicego, & Tistarelli,
2006,2007), fusion of side face and gait (Zhou & Bhanu, 2008), fusion of iris and face
and (Son & Lee, 2005), fusion of palm print and palm vein (Wang, Yau, Suwandy,
& Sung, 2008), fusion of lip and audio (Chetty & Wagner, 2008), etc. Feature level
fusion is also implemented for other felds such as medical image fusion (Kor &
Tiwary, 2004; Patnaik, 2006) object classifcation (Wender & Dietmayer, 2007),
machinery fault diagnosis (Liu, Ma, & Mathew, 2006) and content-based image
retrieval (Rahman, Desai, & Bhattacharya, 2006).
A number of techniques or approaches may be incorporated with feature level
fusion. For example, when feature level fusion is implemented, linear feature
extraction methods are usually employed to reduce the dimension of the resultant
data of the fusion process (Son & Lee, 2005). This is also helpful for reducing the
risk of curse of dimensionality (Rattani, KisKu, Bicego, & Tistarelli, 2006). When
the fused data are in the form of color images, the approach of respectively fusing
data at different channels (Liu & Liu, 2007; Ross & Govindarajan, 2005) is a good
way. In addition, some approaches were used to overcome the problem that mul-
tiple modalities have incompatible feature sets. For example, in order to address
the issue that the minutiae feature point sets of fngerprint are not compatible with
the appearance-based image feature, Rattani, KisKu, Bicego and Tistarelli (2007)
proposed to make the minutiae feature point rotation and translation invariant and
introduced the keypoint descriptor (Ross & Govindarajan, 2005). Feature selection
technique may be involved in feature level fusion (Kumar & Zhang, 2006; Liu,
Meng, Hong, Wang, & Song, 2007).
is prohibited.
13.2 schemes t o f use f eAtures At the f eAture
l evel
Fusion at the feature level can be broadly grouped into two classes. The frst class
covers the direct combination of different trait features obtained by the feature
extraction procedures of different subsystems of the multi-biometric system. The
combined features are then used to perform identifcation or verifcation (as shown in
Figure 13.1). The second class covers feature fusion which involves frst combining
different original biometric traits and then carrying out a certain feature extraction
or feature selection process on the combined biometric traits to obtain the fnal
features for personal authentication (as shown in Figure 13.2). Note that previous
defnitions of fusion at the feature level primarily focus on the frst class of fusion.
However, as shown later in Section 13.3, the second class of fusion is also feasible
and can be adopted by a multi-biometric system. In Section 13.3, the feature fusion
procedure can be viewed as the second class of fusion. This is because after the
face and palm print traits are combined to form x
norm
, the KDCV Algorithm is then
Figure 13.1. Flow chart of the frst class of feature level fusion
Figure 13.2. Flow chart of the second class of feature level fusion
Biometric
traits
Biometric
traits
Feature
Extraction
Feature
Extraction
Matching
Process
Database
Decision
Making
Accept
or reject
n o match
Feature
fusion
Biometric
traits
Biometric
traits
Matching
Process
Database
Decision
Making
Accept

or reject
no match
Feature
fusion
Feature
Extraction
is prohibited.
used to extract features from x
norm
. The feature fusion scheme in Chang, Bowyer
and Sarkar (2003) also belongs to the second class of fusion.
The following sections of this chapter will present two examples of feature level
fusion. The frst example is typical for multi-modal biometrics and clearly shows
the typical fusion process of the frst class of feature level fusion. It frst extracts
features of two biometric traits using the same feature extraction algorithm and then
combines these features to perform personal authentication. The second example
uses multiple presentations of a biometric trait to identify personal identities. This
can be regarded as a typical instance of multi-algorithmic biometrics. Indeed, the
multiple presentations are obtained by applying different algorithms to the same
biometric trait.
13.3 f Ace And PAlm Pr Int f us Ion At the f eAture
l evel
In this section, we present a novel fusion approach to multi-biometrics at the feature
level. We frst combine two kinds of biometrics: the face feature, which is a repre-
sentative contactless biometric trait, and the palm print feature, which is a typical
contacting biometric trait. We perform the Gabor transform on face and palm
print images and combine them at the feature level. The correlation analysis shows
that there is very small correlation between their normalized Gabor-transformed
images. This section also presents a novel classifer, KDCV-RBF, to classify the
fused biometric images. It extracts the image discriminative features using a kernel
discriminative common vectors (KDCV) approach and classifes the features by
using the radial basis function (RBF) network. Two large public face databases (AR
and FERET) and a large palm print database are used to test the fusion approach.
Experimental results demonstrate that the proposed biometric fusion recognition
approach is an effective solution to the small sample size problem.
13.3.1 gabor t ransform of Face and Palm Print images
2D Gabor flters are used mostly in image presentation. Gabor flters are also applied
to texture segmentation, target detection, document analysis, edge detection, etc. A
Gabor flter is basically a Gaussian function modulated by a complex sinusoid. At
the same time, a Gabor flter can be also viewed as a sinusoidal plane of particular
frequency and orientation, modulated by a Gaussian envelope. This can be writ-
ten as h(x,y) = s(x,y)g(x,y), where s(x,y) is a complex sinusoid, and g(x,y) is a 2-D
Gaussian shaped function, known as an envelope. The complex sinusoid is defned
as
2 ( cos sin )
( , )
j ux uy
s x y e
- +
= .
is prohibited.
The 2-D Gaussian function is defned as follows:
2 2
2
2
1
( , )
2
x y
g x y e
+
-
= .
Thus, the Gabor flter is generally defned as follows:
2 2
2
2 ( cos sin )
2
1
( , , , , )
2
x y
j ux uy
G x y u e e
+
-
- +
=
. (13.1)
where u is the frequency of the sinusoidal wave, is the orientation of this flter,
and is the standard deviation of the Gaussian envelop. Note that the fltering result
is a complex number.
The images in the frst row are the resultant images of the Gabor transform
under the conditions of 2. = The images in the second, third and forth rows are
the resultant images of the Gabor transform under the condition of 4 = , 8 = , and
16 = , respectively. For resultant images shown in each row, the orientation of the
Gabor function increases rightwards. That is, the frst image shown in each row is
associated with 0 = while the last image shown in each row is associated with
7
8
=
.
The Gabor function as defned in (13.1) is used as the flter operator. The
size of the face image is 60 60. For the flter parameters, is set as 2, 4, 8, 16,
respectively,
1
u = ,
and is respectively set as
0,
8
,
2
8
,
3
8
,
4
8
,
5
8
,
6
8
and
7
8
.
Thus the Gabor transform used in this section contains four scales and eight ori-
entations. Using the above Gabor function with different parameter values, we can
transform an original image into 32 resultant images. Figure 13.3 shows an original
face image. Figure 13.4 shows the results of Gabor fltering (expressed by magnitude
values). The size of the original palm print image is also 60 60.
is prohibited.
13.3.2 Fusion Procedure
The feature level fusion is implemented using the following procedure(Jing, Yao,
Zhang, Yang, & Li, 2007):
(i) Let X
face
and X
palm
respectively denote sample sets of the face and palm print
images. x
face
is used to denote a sample of X
face
with the size 6060. Performing the
Gabor transform on X
face
, we obtain 32 resultant images (such as the ones shown
in Figure 13.4). Combining them, we get a Gaborface image sample x
Gaborface
, the
initial size of which is 240480. To reduce the computational cost, we downsample
each Gaborface image by a ratio equal to 4. Hence the size of x
Gaborface
is reduced
to 60120. Similarly, for a sample x
palm
from X
palm
, we obtain its Gaborpalm image
x
Gaborpalm
.
(ii) We combine x
Gaborface
and its corresponding x
Gaborpalm
vertically to obtain a
fusion image sample x
fuse
with the size of 120 120. We thus get a fusion image
sample set X
fuse
. Note that the imaging conditions such as lighting conditions and
camera focus settings for the face and palm print images are different. Thus, it is
necessary to perform pixel normalization for fusion images. x
fuse
can be normal-
ized as follows:
Figure 13.3. An original face image

Figure 13.4. Resultant images of Gabor transform on the face image shown in
Figure 13.3

is prohibited.
fuse fuse
norm
fuse
x
x
-
=
,
where
fuse
and
fuse
are the mean and variance of x
fuse
, respectively. We denote the
normalized sample set consisting of x
norm
by X
norm
.
To show the dependency of x
Gaborpalm
on x
Gaborpalm
, we analyze the correlation
between x
Gaborface
and x
Gaborpalm
as follows. Let a and b respectively represent the face
and palm print image parts of x
norm
. We use the following expression to evaluate
their correlation coeffcient
60 120
1 1
60 120 60 120
2 2
1 1 1 1
| ( )( ) |
( , )
( ( ) ).( ( ) )
ij ij
i j
ij ij
i j i j
a a b b
cor a b
a a b b
= =
= = = =
- -
=
- -

,
where
| |
means the absolute value, and , a b are the mean values of a and b,
respectively. We calculate the correlation coeffcients for all samples in X
norm
, and
then obtain the average correlation coeffcient cor(X
norm
). Table 13.1 shows the
correlation analysis of the face and palm print images. For the original image sets
X
palm
and X
face
, we calculate their correlation coeffcients using the same method
presented above. Table 13.1 shows that the correlation of the face and palm print
images is quite small, especially for normalized Gabor-transformed images. This
means that the Gabor transform is helpful in reducing the correlation between dif-
ferent biometric images.
Table 13.1. Correlation coeffcients between original face and palm print images
and correlation coeffcients between resultant images of Gabor transforms of face
and palm print images
The set of samples
Combined Face
and palm print images
Average correlation coeffcient
X
norm
AR and palm print 0.0344
Feret and palm print 0.0395
X
face
and X
palm
AR and palm print 0.1351
Feret and palm print 0.1185
is prohibited.
13.3.3 KDCV-Based Feature Extraction and RBF Classifcation
13.3.3.1 Kernel-Based Nonlinear Discrimination Analysis (Jing, Yao,
Zhang, Yang, & Li, 2007)
For a given nonlinear mapping function , the input data space R
n
can be mapped
into a feature space F. Suppose that the sample set X contains c known pattern
classes, l
i
is the training sample number of the ith class, and there are a total of
1
c
i
i
l l
=
=
training samples. Then ,

b w
S S defned as in (13.1) respectively denote the
between-class scatter matrix and within-class scatter matrix of the feature space.
( )( )
1
1
T
c
b i i
i
S m m m m
l
=
= - -
, ( ) ( ) ( ) ( )
1 1
i
T
c l
i i
w j i j i
i j
S x m x m
= =
= - -

,
(13.1)
where
( )
i
j
x
, j = 1,2,...,l
i
means the jth sample of the ith class,
i
m
is the mean value
of the ith class, and m denotes the mean value of all the training samples. Then
the Fisher criterion in the feature space is as follows:
( )
1
T
b
T
w
S
J
S
=
, (13.3)
For the feature space, the optimal discriminant vector should make
( )
1
J reach
its maximum value, which implies that if we transform the feature space into a new
space by taking as the transforming axis, the new space will have the maximal
linear seperability (Xu, Yang, Lu, & Yu, 2004; Xu, Yang, & Jin, 2003). It has been
proven that a transforming axis can be expressed in terms of
( )
1
l
i i
i
w x
=
=
. (13.4)
We defne a kernel function
( ) ( ) ( )
,
T
i j i j
k x x x x = . (13.5)
Substituting (13.4) and (13.5) into (13.3), we can rewrite (13.3) as
( )
1
( )
T
T
l
KUK
J
K I U K
=
-
, (13.6)
is prohibited.
where K is an l l symmetric matrix with (K)
ij
= k(x
i
,x
j
), I
l
is the l lll identity
matrix, and U is an l l block diagonal matrix. Note that
1 2
( , ,..., )
c
U diag U U U =

where U
i
is a l
i
l
i
matrix, the elements of which are all equal to

1
i
l
.
We defne
'
b
S KUK = and
'
( )
w l
S K I U K = - and respectively call them between-class
matrix and within-class scatter matrix of the kernel space implicitly defned by the
kernel function.
13.3.3.2 Feature Extraction Using KDCV Algorithm
The DCV (discriminative common vectors) algorithm aims at obtaining the optimal
discriminant matrix W consisting of a number of discriminant vectors in the null
space of the within-class scatter matrix of the original feature space. It is based on
the following expression:
| | 0
( ) argmax | |
T
w
T
b
W S W
J W W S W
=
= , (13.7)
where S
b
means the between-class scatter matrix of the original feature space. Two
steps are needed for implementing the DCV algorithm. The frst step is to obtain
the null space of S
w
. The second step is to get the optimal discriminant matrix
W and the DCV. Using the DCV algorithm, we implement KDCV-based feature
extraction as follows:
Step 1 Calculate the common vectors.
Let R
d
denote the original sample space, V denote the non-null space of
'
w
S ,
and V
denote the null space of S

w
. Thus we obtain the following expressions:
'
{ | 0, 1, 2,..., }
k w k
V span S k r = = and
'
{ | 0, 1, 2,..., }
k w k
V span S k r r l
= = = + + ,
where r is the rank of S
w
. Indeed,
1 2
. . .
r
are the eigenvectors cor-
responding to the nonzero eigenvalues of S
w
, whereas
1 2
. . .
r r l + +
are the
eigenvectors corresponding to the zero eigenvalues of
'
w
S . Let
1 2
[ . . . ]
r
Q =
and
1 2
[ . . . ]
r r l
Q
+ +
= . Because
d
R V V
= , every sample
( )
i
m
x
can be
expressed as follows:
( ) ( ) ( )
i i i
m m m
x y z = +
, (13.8)
where ( ) ( )
i T i
m m
y QQ x = , ( ) ( )
T
i i
m m
z QQ x = . ( ), ( )
i i
m m
y z respectively represent
the difference vector and common vector parts of ( )
i
m
x . It has been proved that
all samples of the ith class have the same common vectors. Hereafter let ( )
i
com
x
is prohibited.
represent the common vector of the ith class instead of ( )
i
m
z . Then we rewrite Eq.
(13.8) and calculate
( )
i
com
x
as follows:
( ) ( ) ( ) ( ) ( )
i i i i T i
com m m m m
x x y x QQ x = - = - . (13.9)
We therefore obtain a common vectors set
1 2
{ ( ), ( ),..., ( )}
c
com com com
A x x x = .
Step 2 Calculate the optimal discriminant vector
com
S is used to represent the total scatter matrix of A, i.e.
( )( )
1
( ) ( )
T
c
i i
com com com com com
i
S x x
=
= - -
, (13.10)
where
1
1
( )
c
i
com com
i
x
c
=
=

. The transform matrix W consisting of the optimal
discriminant vectors will satisfy the following expression:
'
| | 0
( ) argmax | |
T
w
T
com
W S W
J W W S W
=
= , (13.11)
where
'
w
S means the within-class matrix associated with
com
S .
Obviously, W should be composed of the eigenvectors corresponding to the
nonzero eigenvalues of
com
S . Generally, all classes are independent of each other.
Thus, all common vectors are also independent of each other and the rank of
com
S
is c 1. We can obtain nonlinear DCVs using
( ), 1, 2,...,
T i
i com
y W x i c = = . (13.12)
Similar to DCV, y
i
is also identical for the ith class and the feature dimension of
y
i
is c 1. For any testing sample ( )
test
x , we transform it into a feature vector y
test

using
( )
T
test test
y W x =
. The procedure presented above is called the discriminative
feature extraction procedure.
13.3.3.3 Classifcation Using RBF
The RBF network has been widely applied to regression analysis and pattern rec-
ognition. A typical RBF network contains the following three layers: input layer,
hidden layer and output layer. The neuron number of the input layer is decided by
the feature vector dimension of samples, while the number of neurons in the hidden
layer is adjustable. The neurons of the output layer are as many as the pattern classes.
We set expected output values for the ith class as follows: the expected output value
of the ith neuron is set to 1 and the expected output values of the other neurons
are set to 0. The weights between the input and hidden layers are fxed at 1. The
is prohibited.
RBF network determines the classes of KDCV features, which are the derivation
of the feature level fusion images. Note that a Gaussian kernel function serves as
the kernel function for both the KDCV feature extraction procedure and the RBF
network. Since the RBF network is used only for classifying KDCV features, we
hereafter call it the KDCV-RBF classi. er. Figure 13.5 briefy describes the main
steps of feature extraction and classifcation:
Step 1 The face and palm print image sample sets X
face
and X
palm
are processed
by the feature level fusion approach to get a fusion image sample X
norm
. X
norm
is
a one-dimensional vector and the corresponding one-dimensional sample set is
denoted by X.
Step 2 Based on X, we obtain the transform matrix W and c nonlinear DCVs using
the KDCV algorithm. These nonlinear DCVs are used to construct the training sample
set Y
train
. After all testing samples are transformed by W, the testing sample set Y
test
is
obtained. The Gaussian kernel function is defned as
2
1 2 1 2
( , ) exp( || || 2 ) k x x x x =
where is set as the variance of the training sample set of X.
Step 3 Respectively input Y
train
and Y
test
into the RBF network. The input layer,
hidden layer and output layer have c 1, c, and c neurons, respectively. All the c
nonlinear DCVs are taken as the clustering centers and stored in the hidden neurons.
The Gaussian radial basis function used here is
2
1 2 1 2
( , ) exp( || || 2 ) k x x x x = with
the parameter value 10 = . Note that the weights between the input and hidden
layers are all 1. It is necessary to calculate only the weights matrix W
n
between
the hidden and output layers. When Y
train
is inputted into the network, G is used to
denote the hidden layer values, WRBFG means the actual output values, and F is
the expected output values. W
n
is calculated using
Figure 13.5. Multi-modal biometric recognition procedure
Nonlinear
feature
extraction
Gaborface
Gaborpalm
Pixel
level

Recognition
result
KDCV-RBF
classifier
is prohibited.
n
W G F
+
= , (13.13)
where G
+
denotes the pseudoinverse of G. After this, the trained RBF network is
used to classify the sample set Y
test
.
13.3.3.4 Experimental Results
In this section, two face databases (AR and FERET) and a palm print database are
used to test the proposed approach. For our experiments on biometric fusion, the
AR database (please refer to Section 3.4.4) and the FERET database are respectively
paired with the palm print database to form two pairs of databases. The experiments
compare feature level and classifer recognition approaches. Feature level approaches
will use different image data and classify these data based on the KDCVRBF
classifer (here abbreviated to KDRC). Classifer approaches will use the fused
Gabor-transformed images and classify them using different feature extraction
and classifcation approaches. We implement several feature level approaches as
follows: (1) Using single-mode original images respectively from the AR, FERET
and palm print databases, we carry out in sequence AR-KDRC, FERET-KDRC
and Palm-KDRC. (2) Instead of carrying out a Gabor transform, we use directly
original images to perform feature level fusion of the two pairs of databases (i.e.
AR and palm print, and FERET and palm print) and then classify the fusion results
by using the RBF network. The corresponding implementation processes based on
the above two pairs of databases are called ARPalm-Originalfusion-KDRC and
FERETPalm-Originalfusion-KDRC, respectively. (3) We perform Gabor transform
and feature level fusion in sequence. Two corresponding implementation processes
based on the two pairs of databases are respectively called ARPalm-Gaborfusion-
KDRC and FERETPalm-Gaborfusion-KDRC.
Several classifer approaches are implemented as follows: (1) We carry out AR-
Palm-Gaborfusion-KDNC and FERETPalm-Gaborfusion-KDNC for two pairs of
databases using the KDCV and NN classifer (abbreviated to KDNC). (2) Employing
the DCV and NN classifer (DNC), we implement ARPalm-Gaborfusion-DNC
and FERETPalm-Gaborfusion-DNC. (3) Using the Kernel PCA and NN classifer
(KPNC), we carry out ARPalm-Gaborfusion-KPNC and FERETPalm-Gaborfusion-
KPNC. (4) Using the KDRC classifer, we implement ARPalm-Gaborfusion-KDRC
and FERETPalm-Gaborfusion-KDRC which will be presented latter.
To classify the sample features obtained using the DCV algorithm (Cevikalp.
Neamtu, Wilkes, & Barkana, 2005), we use the NN (nearest neighbor) classifca-
tion based on the Euclidean distance. In implementing the Kernel PCA (KPCA)
(Scholkopf, Smola, & Muller, 1998), we use the same Gaussian kernel function as
the one used in KDCV feature extraction.
is prohibited.
For the AR face database, the images of the 119 individuals participating in
the two image capturing sessions were selected and used in our experiment for
a total number of 3094 (=11926) samples. Every image is resized to be 6060.
The FERET database employed in our experiment includes 2200 facial images
generated from 200 individuals (each having 11 images) (Phillips, Wechsler, Huang,
& Rauss,1998). They are the images named with two-character strings from ba
to bk. We performed a histogram equalization for the cropped images and scaled
them to 6060.
The palm print database used in the experiments is from the Hong Kong Poly-
technic University. This database contains palm print images of 189 individuals.
The subjects are mainly volunteers of the Hong Kong Polytechnic University. Each
person provided 20 images and the obtained database has a total of 3780 (=18920)
images. For each palm of every subject, the number of the acquired images is 10.
Each of original images contained 384 284 pixels and had a resolution of 75 dpi.
We obtained the subimages with a fxed size 128128 from the original images by
using the image processing approach in (Zhang, Kong, You, & Wong, 2003). In order
to reduce the computational cost, we reduced the size of each subimage to 6060
pixels. We took these subimages as palm print image samples for the experiments.
The palm was imaged under varying conditions including unfxed illumination,
shift and rotation of the palm. In addition, the image may also be slightly affected
by the way the hand is posed, shrunk, or stretched. Figure 13.6 shows the twenty
images of one subject from the palm print database.
Figure 13.6. The twenty images of one subject from the palm print database used
in this section
(a) (b) (c) (d) (e) (f) (g)
(h) (i) (j) (k) ( l) (m) (n)
(o) (p) (q) (r) (s) (t)
is prohibited.
13.3.3.4.1 Experiment on AR Faces and Palm Print
Note that the palm print database contains 189 classes each containing 20 samples
and the AR face database has 119 classes each having 26 samples. We used the frst
20 samples of each subject of the AR face database and all the pamlprint images
of each of the frst 119 subjects of the palm print database. We set the numbers of
training samples per class to be 2 and 3, respectively, and took the remainder as
(a)
(b)
Figure 13.7. (Jing, Yao, Zhang, Yang, & Li, 2007) Random testing results on AR
and palm print databases in the case where the number of training samples per
class is 2
is prohibited.
testing samples. The results of 30 random tests of different methods on two and
three training samples per class will be shown in Figure 13.7 and Figure 13.8,
respectively. Firstly, we set the number of training samples per class to be two so
that here are 238 (=1192) training samples and 2142 (=11918) testing samples.
From 13.7(a), we can see that the feature level fusion approaches clearly improve
the recognition performance of AR-KDRC and Palm-KDRC, and ARPalm-
Gaborfusion-KDRC performs best. From Figure 13.7(b), we can see that in most
cases, ARPalm-Gaborfusion-KDNC outperforms ARPalm-Gaborfusion-DNC and
Figure 13.8. (Jing, Yao, Zhang, Yang, & Li, 2007) Random testing results on AR
and palm print databases in the case where the number of training samples per
class is 3
(a)
(b)
is prohibited.
ARPalm-Gaborfusion-KPNC. In addition, ARPalm-Gaborfusion-KDRC obtains
better recognition results than ARPalm-Gaborfusion-KDNC. Secondly, we set the
number of training samples per class to be 3 so that there are 357 (=119 3) training
samples and 2023 (=11917) testing samples. Figure 13.8(a) shows that ARPalm-
Gaborfusion-KDRC gets the best results among all the approaches. Figure 13.8(b)
shows that ARPalm-Gaborfusion-KDRC performs best in almost all cases. Table
13.2 shows recognition results of all the approaches. Using a single biometric trait,
AR-KDRC and Palm-KDRC obtain total average recognition rates of 71.28% and
63.81%, respectively. ARPalm-Gaborfusion-KDRC improves the total rate to 94.40%
by exploiting bimodal-biometric traits and Gabor transform. Table 13.2 also shows
that KDCV outperforms DCV and KPCA in extracting features of biometric traits,
and that the RBF network is more suitable for classifying the KDCV features than
the NN classifcation. The total average recognition rates of the KDRC classifer is
signifcantly better than those of both the DNC classifer at 7.46% (=94.40%86.94%)
and the KPNC classifer at 11.04% (=94.40% 83.36%).
13.3.3.4.2 Experiment on FERET Faces and Palm Print
Note that the FERET face database contains 200 subjects each having 11 samples
and the palm print database contains 189 subjects and each subject has 20 samples.
To implement fusion, we use all the 11 samples of each class of the frst 189 face
classes, and the frst 11 samples of each class of all the 189 palm print classes. We
Table 13.2. Recognition results of different approaches on AR and palm print da-
tabases
Approaches
Number of training samples per class
Total mean
values
2 3
The average
recognition rate (%)and
the variance(%)
The average
recognition rate(%)
and the variance(%)
AR-KDRC 65.67 (13.16) 76.88 (12.80) 71.28
Palm-KDRC 63.33 (1.30) 64.29 (1.57) 63.81
ARPalm-Originalfusion-KDRC 76.67 (11.68) 85.70 (9.57) 81.19
ARPalm-Gaborfusion-KDNC 89.12 (2.87) 91.35 (3.00) 90.24
ARPalm-Gaborfusion-DNC 86.54 (2.61) 87.33 (3.39) 86.94
ARPalm-Gaborfusion-KPNC 80.70 (7.94) 86.02 (6.79) 83.36
ARPalm-Gaborfusion-KDRC 92.66 (3.08) 96.14 (3.53) 94.40
is prohibited.
will set the numbers of training samples per class to be two and three, respec-
tively, and the remainder is used as testing samples. When the number of training
samples per class is set to be two, there are 378 (=189 2) training samples and
1701 (=189 9) testing samples in total. Figure 13.9(a) shows that feature level
fusion approaches are superior to single biometric approaches, and FERETPalm-
Gaborfusion-KDRC performs better than FERETPalm-Originalfusion-KDRC.
Figure 13.9(b) shows that in most cases, FERETPalm-Gaborfusion-KDNC outper-
forms FERETPalm-Gaborfusion-DNC and FERETPalm-Gaborfusion-KPNC. In
addition, FERETPalm-Gaborfusion-KDRC obtains better recognition results than
(a)
(b)
Figure 13.9. Random testing results on the FERET and palm print databases in the
case where 2 training samples per class
is prohibited.
FERETPalm-Gaborfusion-KDNC in all the cases. When the number of training
samples per class is set to be three, there are 567 (=189 3) training samples and
1512 (=189 8) testing samples in total. Figure 13.10(a) shows that for random tests
of three training samples per class, the feature level fusion approaches outperform
FERET-KDRC and Palm-KDRC, and FERETPalm-Gaborfusion-KDRC obtains
the best recognition results. Figure 13.10(b) shows that for tests of three training
samples per class, FERETPalm-Gaborfusion-KDRC performs best. Table 13.3 shows
comparison of the average recognition results of all the approaches. Using single
biometric trait, FERET-KDRC and Palm-KDRC obtain total average recognition
Figure 13.10. Random testing results on FERET and palm print databases in the
case where 3 training samples per class
(a)
(b)
is prohibited.
rates of 63.35% and 57.95%, respectively. FERETPalm-Gaborfusion-KDRC im-
proves the total rate to 91.22% by using the Gabor-based feature level fusion. Table
13.3 also shows that KDCV outperforms DCV and KPCA, and the RBF network
is more suitable for classifying the KDCV features than the NN method. The total
average recognition rate of the KDRC classifer is respectively 13.64% (91.22%
77.58%=13.64%) higher and 8.95% (91.22% 82.27%=8.95%) higher than those
of both the DNC and KPNC classifers.
13.3.3.4.3 Analysis of Experimental Results
Based on Tables 13.2 and 13.3, Figure 13.11 shows the total average recognition
rates of the two groups of experiments presented above. From Figure 13.11, we can
see that the total recognition rate rises from 67.32% (face recognition) and 60.88%
(palm print recognition) to 92.81% (multi-modal fusion biometrics). Based on
Gabor transform of two biometric traits, FacePalm-Gaborfusion-KDRC is 12.26%
better than FacePalm-Originalfusion-KDRC, which uses the original biometric
images. The proposed KDRC classifer improves the total recognition rates of
two representative classifers (DNC and KPNC) by 10.55% and 9.99%. These
experimental results can be evaluated by using the null hypothesis statistical test
based on the Bernoulli model (Beveridge, She, Draper, & Givens, 2001; Yambor,
Draper, & Beveridge, 2002), which tells us that if the resulting p-value is below the
desired signifcance level (i.e. 0.01), the null hypothesis should be rejected and the
Table 13.3. Average recognition results of the FERET and palm print databases
Methods
Number of training samples per class
2 3
Total mean
values
Mean values and
variances (%)
Mean values and
variances (%)
FERET-KDRC 58.11 7.79 68.58 7.19 63.35
Palm-KDRC 58.25 4.06 57.65 4.54 57.95
FERETPalm-Originalfusion-KDRC 76.70 6.40 83.11 4.89 79.91
FERETPalm-Gaborfusion-KDNC 83.93 1.72 85.77 1.17 84.85
FERETPalm-Gaborfusion-DNC 76.89 6.09 78.26 7.24 77.58
FERETPalm-Gaborfusion-KPNC 81.40 4.55 83.14 4.30 82.27
FERETPalm-Gaborfusion-KDRC 89.90 3.32 92.53 2.46 91.22
is prohibited.
performance difference between two algorithms are considered to be statistically
signifcant. Table 13.4 shows the statistical analysis of the recognition performance
difference between the proposed FacePalm-Gaborfusion-KDRC approach and Face-
Palm-Gaborfusion-KDNC. In Table 13.4, the item Number of signifcant differ-
ences (p < 0.01) means the number of statistically signifcant differences between
the two approaches occurring at a signifcance level of p < 0.01. It seems that in
most trials, there are signifcant differences between FacePalm-Gaborfusion-KDRC
and FacePalm-Gaborfusion-KDNC. Table 13.5 shows the total statistical difference
analysis for FacePalm-Gaborfusion-KDRC and other methods. The item Number
Table 13.4. Statistical analysis of the recognition performance difference of Face-
Palm-Gaborfusion-KDRC and FacePalm-Gaborfusion-KDNC
Involved databases
Number of training
samples per class
Number of signifcant
differences
AR and Palm print
2 27
3 30
FERET and Palm print
2 28
3 30
Figure 13.11. (Jing, Yao, Zhang, Yang, & Li, 2007). Total average recognition rates
of face and palm print biometric recognition
0
10
20
30
40
50
60
70
80
90
100
Face-KDRC
Palm-KDRC
FacePalm-Originalfusion-KDRC
FacePalm-Gaborfusion-KDNC
FacePalm-Gaborfusion-DNC
FacePalm-Gaborfusion-KPNC
FacePalm-Gaborfusion-KDRC
is prohibited.
of signifcant differences (p < 0.01) for random test (60 times) means the number
of statistically signifcant differences existing between the two approaches while
the 60 random tests, each being associated with two or three training samples per
class, are conducted. Table 13.5 shows that the recognition performance of Face-
Palm-Gaborfusion-KDRC has a signifcant difference in comparison with those
of the other methods.
13.3.3.4.4 Conclusion
The proposed feature level biometric fusion approach frst combines the normalized
Gabor-transformed face and palm print images at the feature level and then uses
KDCV-RBF (KDRC) to classify the fused biometric images. The experimental re-
Other approaches Involved databases
Number of
signifcant differ-
ences (p <0.01)
for 60 random
tests
Average number
of signifcant
differences
Face-KDRC
AR and Palm print
60
60
60
Palm-KDRC
AR and Palm print
60
60
60
FacePalm-Originalfusion-
KDRC
AR and Palm print
60
60
60
FacePalm-Gaborfusion-
KDNC
AR and Palm print
57
58
57.5
DNC
AR and Palm print
55
59
57
KPNC
AR and Palm print
60
60
60
Table 13.5. The total performance difference between FacePalm-Gaborfusion-KDRC
and other approaches
is prohibited.
sults on two large face databases (AR and FERET) and a large palm print database
demonstrate that this feature level biometric fusion approach is an effective solution
to the small sample biometric recognition problem. While this section provides with
us a typical example of bimodal biometrics, it also allows primary characteristics
of the frst class of feature level fusion to be clearly shown.
13.4 f us Ion of mul t IPle f eAture Present At Ions
A real-world object can be presented in different ways. For example, it can be
described in terms of its shape or color. Different feature extraction procedures
may also have different focuses and produce different descriptions. For example,
some feature extraction procedures may emphasize holistic characteristics whereas
some other procedures may emphasize local characteristics. This makes it possible
that the feature extraction results of different feature extraction procedures are
complementary. In this book the term multiple feature presentation is used to
denote multiple feature extraction results of a same biometric trait obtained using
different feature extraction procedures. This section describes a technique that frst
takes phase and orientation information generated from the palm print image as
multiple feature presentations of the palm print and then integrates these feature
presentations to perform personal authentication.
13.4.1 Extraction of Phase and Orientation information
Personal authentication presented in this section is based on a multiple feature
extraction approach to a biometric trait. This approach is mainly involved in the
fusion of the phase and orientation information of the 2D Gabor transform of the
biometric image. While the proposed approach extracts the phase information (called
PhaseCode) of an image by using four 2-D Gabor flters with different orientations,
the orientation information (called OrientationCode) of the image is also extracted.
The PhaseCode and the OrientationCode are fused to make a new feature, called
the Phase and Orientation Code (POC). The POC-based palm print authentication
approach proposed in this paper is called the POC approach. At the matching stage,
a modifed Hamming distance is used to measure the similarity of two POCs.
The POC is calculated by using a 2D Gabor flter as defned in (13.1).
The imaginary part of the result of the Gabor transform inherently has no DC
(direct current) because of the odd symmetry, whereas the real part has DC. How-
ever, the real part can be normalized to zero DC as follows:
is prohibited.
2
( , , , , )
( , , , , ) ( , , , , )
(2 1)
n n
i n k n
G x y u
G x y u G x y
n
-
=- =-
= -
+

, (13.14)
where (2n = 1)
2
is the size of the flter. After the normalization process, these flters
have no DC response. This is very helpful for reducing the infuence on the fltered
results of the strength of illumination. We use four Gabor flters whose orientations
are 0
0
, 45
0
, 90
0
, 135
0
to extract the POC. These flters are respectively denoted by
G
j
, j = 0,1,2,3.
Let I stand for a palm print image and suppose that I has been fltered by G
j
( j
= 0,1,2,3) as follows:
j
G j
I I G = , (13.15)
where * denotes the convolution operation. The magnitude of the fltering result
is formulated as
( , ) ( , ) ( , )
j j
j G G
M x y I x y I x y
-
= , (13.16)
where - denotes complex conjugate. The orientation of the point (x,y) is calculated
using the following expression:
0,1,2,3
( , ) arg max ( , )
j
j
O x y M x y
=
=
. (13.17)
O(x,y) is called the OrientationCode of the point (x,y). The PhaseCode can be
calculated (Kong, Zhang, & Li, 2003) using:
( , )
1, Re( ( , )) 0;
( , )
0,
j
G x y
R
if I x y
P x y
otherwise

( , )
1, Im( ( , )) 0;
( , )
0,
j
G x y
I
if I x y
P x y
otherwise

(13.18)
( , )
1, Im( ( , )) 0;
( , )
0,
j
G x y
I
if I x y
P x y
otherwise

(13.19)
P = (P
R
,P
I
,O) is called Phase Orientation Code (POC). Figure 13.12 shows POCs
of three palm print images.
is prohibited.
13.4.2 Similarity Measure Based on Modifed hamming
distance
We now will show that how the multiple feature representations i.e. POCs are
converted into similarity measures of palm print images by using the Hamming
distance. The Hamming distance is the number of positions in two strings of
equal length for which the corresponding elements are different. For example, the
Hamming distance between strings 2143896 and 2233796 is 3 and the Hamming
distance between 1011101 and 1001001 is 2. The Hamming distance can be used
to measure the similarity between two strings. For example, because each string
of the frst couple of strings (2143896 and 2233796) are 7 characters in length and
there are three different elements, the percentage (7 - 3) / 7 = 57% can be used to
represent the similarity between the two strings. For the second couple of strings
(1011101 and 1001001), the similarity is (7 - 2) / 7 = 71%.
Here a modifed Hamming distance is defned to measure the similarity of two
binary matrices. A binary matrix is often used to represent an image containing
an object. Because of imperfect preprocessing, there may exist some non-object
pixels in the preprocessed images. Let I
1
and I
2
denote two images and suppose
Figure 13.12. POCs of three palm print images: (a),(b), (c) denote the original palm
prints; (d),(e),(f) show the real parts of the POCs; (g),(h),(i) represent the imaginary
parts of the POC; (j),(k),(l) denote the OrintationCode of the POC.
is prohibited.
that M
1
and M
2
are the corresponding masks (Zhang, Kong, You, & Wong, 2003)
that identify the location of the non-palm-print pixels of the two palm print images.
respectively. Then the valid pixels used for matching can be represented by a binary
matrix M:
1 2
( , ) ( , ) ( , ) M i j M i j M i j =
. (13.20)
where means the logical AND operator.
Let P
R
and P
I
denote the PhaseCode of I
1
. Let O
1
denote the OrientationCode of
I
1
. Q
R
,Q
I
, and O
2
mean the PhaseCode and the OrientationCode of I
2
, respectively.
Then C
1
= (P
R
,P
I
,O
1
), C
2
= (Q
R
,Q
I
,O
2
) can be used to respectively represent I
1
and
I
2
. The binary matrixes H
R
and H
I
, which respectively stand for the differences
between P
R
and Q
R
, and between P
I
and Q
I
, are computed as below:
( , ) ( , ) ( , )
R R R
H i j P i j XOR Q i j =
. (13.21)
( , ) ( , ) ( , )
I I I
H i j P i j XOR Q i j = . (13.22)
The difference between O
1
and O
2
is expressed in terms of the following binary
matrix H
O
:
If O
1
= O
2
, then H
O
(i,j) = 1;
If O
1
= O
2
, then H
O
(i,j) = 0;
The following modifed Hamming distance is used to measure the similarity
D(I
1,
I
2
) between I
1
and I
2
:
1 1
1 2
1 1
1 1
1 1
( , ) ( ( , ) ( , ))
( , )
2 ( , )
( , ) ( ( , ) ( , ))
2 ( , )
N N
R O
i j
N N
i j
N N
I O
i j
N N
i j
M i j H i j H i j
D I I
M i j
M i j H i j H i j
M i j
= =
= =
= =
= =

=

+

. (13.23)
where N N is the size of the images, and mean the logical AND and
OR, respectively. Note that D(I
1
,I
2
) is between 0 and 1 and a large D(I
1
,I
2
) value
means that I
1
and I
2
have a great similarity.
is prohibited.
13.4.3 experimental r esults
We tested the proposed palm print authentication approach using a database con-
taining 7,605 palm prints collected from 392 different palms. These images were
taken from the people of different ages and both sexes in the Hongkong Polytechnic
University. These palm print images were captured twice, at an interval of around
two months. The images in the database are of two different sizes, 384 284 and
768 568. The images of size 768 568 were resized to 384 284 pixels. Then
the central 128 128 pixels of every obtained image were cropped to extract the
POC using the preprocessing technique described in (Zhang, Kong, You, & Wong,
2003).
We performed verifcation experiment as follows: every sample in the database
was matched against each of the remaining ones. If two palm print images from
an identical palm are matched correctly, the matching result is called genuine
matching; otherwise, the matching result is called impostor matching. A total of
28,914,210 matchings have been performed, of which 70,502 matching results are
genuine matchings. Figure 13.13 shows the distance distributions of the genuine
matching and impostor matching. It is clear that two distinct peaks are respectively
associated with the two distance distributions. The peak of the distance distribu-
tion of the genuine matching is located at about 0.41, while the peak of the distance
distribution of the impostor matching is located at about 0.84. Since the two peaks
are widely separated and the two curves intersect very little, the Poc approach is
able to authenticate palm prints with a high accuracy.
Figure 13.13. Distributions of the genuine and imposter matching distances
is prohibited.
Figure 13.14 presents the receiver operating characteristic (ROC) curve of the
Poc approach. Experiments of the Sobel algorithm (Han, Chen, Lin, & Fan, 2003)
and the improved PalmCode (Zhang, Kong, You, & Wong, 2003) on this database,
i.e. the PhaseCode method (Kong & Zhang, 2004), were also implemented and
their ROC curves are also shown in Figure 13.14. The equal error rates (EERs) of
each approach is listed in Table 13.6. According to Figure 13.14 and Table 13.6, we
know that the Poc approach and PhaseCode method perform much better than the
Sobel algorithm. The EERs of the Poc approach and PhaseCode method are 0.31%
and 0.56%, respectively. From Figure 13.15 which shows the ROC curves of the
Poc approach and PhaseCode method with the log scale, we can see that when the
FRR is 2%, the FARs of the Poc approach and the PhaseCode method are about
0.0036% and 0.0078%, respectively.
Figure 13.14. ROC curves and the ERR of the Sobel algorithm, the PhaseCode
method and the POC approach with the log scale
Table 13.6. Matching distances between two POCs on the palm print images (a),(b),
(c) in Figure 13.12
(a) (b) (c)
(a) 0 0.8339 0.8226
(b) 0.8339 0 0.7989
(c) 0.8226 0.7989 0
is prohibited.
13.4.4 conclusion
The POC approach is an improvement of the PhaseCode method. The POC approach
uses four 2D Gabor flters to calculate the phase and the orientation information,
which are then combined to obtain a new feature, i.e. phase and orientation code
(POC). Palm print identifcation or verifcation can be performed by the Poc approach
based on a modifed Hamming distance. The experimental results on a palm print
database show that the Poc approach is powerful for palm print authentication. The
Poc approach and the PalmCode approach perform much better than the Sobel-based
approach. The EER of the Poc approach decreased from 0.56% of the PhaseCode
method to 0.31% and when the FRR is 2%, the FARs of the POC approach has
decreased from 0.0078% of the PhaseCode method to 0.0036%.
Figure 13.15. ROC curves of the POC approach and the PhaseCode method with
the log scale
Table 13.7. The EERs of the Sobel algorithm, the PhaseCode method and the POC
approach
Approach POC Sobel Algorithm PhaseCode
ERR value (%) 0.31 5.0 0.56
is prohibited.
13.5 comments
Among the three classes of information fusion, feature level fusion is able to
convey the richest information. However, the fact there are only a little number of
literatures on feature level fusion shows that fusion at the feature level has probably
not received the amount of attention it deserves. Feature level fusion may encounter
two potential problems. First, it might be diffcult to perform a combination at the
feature level when the relationship between the feature spaces of different biometric
traits is unknown and the feature representations are not compatible. Second, it is
possible that the feature values of a certain individual biometric trait may not be
accessible because of their proprietary nature or because the user does not have
this biometric trait.
r eferences
Beveridge, J. R., She, K., Draper, B., & Givens, G. H. (2001). Parametric and non-
parametric methods for the statistical evaluation of human ID algorithms. Proc. of
the Third Workshop Empirical Evaluation of Computer Vision Systems.
Cevikalp, H., Neamtu, M., Wilkes, M., & Barkana, A. (2005). Discriminative
common vectors for face recognition. IEEE Trans. Pattern Anal. Mach. Intell.,
27(1), 4-13.
Chang, K., Bowyer, K. W., & Sarkar, S. (2003). Comparison and combination of
ear and face images in appearance-based biometrics. IEEE Trans. Pattern Anal.
Machine Intell., 25(9), 1600-1165.
Chetty, G., & Wagner, M. (2008). Robust face-voice based speaker identity verifca-
tion using multilevel fusion. Image and Vision Computing, 26, 1249-1260.
Choi, K., Choi, H., & Kim, J. (2005). Fingerprint mosaicking by rolling and sliding.
Proceedings of Audio- and Video-based Biometric Person Authentication (AVBPA),
Rye Brook, NY (pp. 260269).
Gunatilaka, A. H., & Baertlein, B. A. (2001). Feature-level and decision-level fusion
of noncoincidently sampled sensors for land mine detection. IEEE Trans. Pattern
Gunes, H., & Piccardi, M. (2005). Affect recognition from face and body: early
fusion vs. late fusion. Proc. of 2005 IEEE International Conference on Systems,
Man and Cybernetics (pp. 3437-3443).
is prohibited.
Han, C., Chen, H., Lin, C., & Fan, K. (2003). Personal authentication using palm-
print features. Pattern Recognition, 36(2), 371-381.
Jain, A. K., Ross, A., & Prabhakar, S. (2004). An introduction to biometric Recog-
nition. IEEE Trans. Circuits and Systems for Video Technology, Special Issue on
Image- and Video-Based Biometrics, 14(1), 4-20.
Jing, X. Y., Yao, Y. F., Zhang, D., Yang, J. Y., & Li, M. (2007). Face and palmprint
pixel level fusion and Kernel DCV-RBF classifer for small sample biometric rec-
ognition. Pattern Recognition, 40, 3209-3224.
Kober, R., Harz, U., & Schiffers, J. (1997). Fusion of visual and acoustic signals
for command-word recognition. Proc. of International Conference on Acoustics,
Speech and Signal Processing (pp. 14951497).
Kong, W., & Zhang, D. (2004). Feature-level fusion for effective palmprint au-
thentication. Proc. of International Conference on Biometric Authentication (pp.
761-767).
level fusion. Pattern Recognition, 39, 478-487.
Kong, W. K., Zhang, D., & Li W. (2003). Palmprint feature extraction using 2-D
Gabor flters. Pattern Recognition, 36(10), 2339-2347.
Kor, S., & Tiwary, U. (2004). Feature level fusion of multimodal medical images
in lifting wavelet transform domain. 26th Annual International Conference of the
IEEE Engineering in Medicine and Biology Society, (IEMBS 04),1, 1479 -1482.
Kumar, A., & Zhang, D. (2006). Personal recognition using hand shape and texture.
IEEE Transactions on Image Processing, 15(8), 2454 2461.
Liu, X., Ma, L., & Mathew, J. (2006). Machinery fault diagnosis based on feature
level fuzzy integral data fusion techniques. 2006 IEEE International Conference
on Industrial Informatics (pp. 857862).
Liu, Z., & Liu, C. (2007). Fusion of the complementary Discrete Cosine Features in
the YIQ color space for face recognition. Computer Vision and Image Understand-
ing, doi:10.1016/j.cviu.2007.12.002.
Liu, W., Meng, H., Hong, W., Wang, L., & Song, J. (2007). A new method for di-
mensionality reduction based on multivariate feature fusion. IEEE International
Conference on Integration Technology, 2007 (ICIT 07) (pp. 108111).
is prohibited.
Patnaik, D. (2006). Biomedical image fusion using wavelet transforms and SOFM
neural network. IEEE International Conference on Industrial Technology 2006
(ICIT2006) (pp.11891194).
Phillips, P. J., Wechsler, H., Huang, J., & Rauss, P. (1998). The FERET database
and evaluation procedure for face-recognition algorithms. Image Vision Comput.,
16(5), 295-306.
Rahman, M. M., Desai, B. C., & Bhattacharya, P. (2006). A feature level fusion
in similarity matching to content-based image retrieval. Proceedings of the 9th
International Conference on Information Fusion (pp. 10-13).
Ratha, N., Connell, J., & Bolle, R. (1998). Image mosaicing for rolled fngerprint
construction. Proceedings of 14th International Conference on Pattern Recogni-
tion, Brisbane (pp. 16511653).
Rattani, A., KisKu, D. R., Bicego, M., & Tistarelli, M. (2006). Robust feature-level
multibiometric classifcation. 2006 Biometrics Symposium: Special Session on
Research at the Biometric Consortium Conference (pp.16).
Rattani, A., Kisku, D. R., Bicego, M., & Tistarelli, M. (2007). Feature level fusion of
face and fngerprint biometrics. First IEEE International Conference on Biometrics:
Theory, Applications, and Systems (BTAS 2007) (pp.16).
biometrics. Proc. of SPIE Conference on Biometric Technology for Human Iden-
tifcation II (pp. 196-204).
Ross, A., & Jain, A. (2003). Information fusion in biometrics. Pattern Recognition
Letters, 24(13), 2115-2125.
Scholkopf, B., Smola, A,. & Muller, K. (1998). Nonlinear component analysis as a
Kernel eigenvalue problem. Neural Comput., 10(5), 1299-1319.
ligence, 29(4), 687-700.
Singh, R., Vatsa, M., Ross, A., & Noore, A. (2005). Performance enhancement of 2d
face recognition via mosaicing. Proceedings of Fourth IEEE Workshop on Automatic
Identifcation Advanced Technologies (AuotID), Buffalo, USA (pp. 6368).
Son, B., & Lee, Y. (2005). Biometric authentication system using reduced joint
feature vector of iris and face. Lecture Notes in Computer Science, 3546/2005,
1611-3349.
is prohibited.
Wang, J., Yau, W., Suwandy, A., & Sung, E. (2008). Person recognition by fusing
palmprint and palm vein images based on Laplacianpalm representation. Pattern
Recognition, 41 (2008), 1514 1527.
Wender, S., & Dietmayer, K. C. J. (2007). A feature level fusion approach for object
classifcation. 2007 IEEE Intelligent Vehicles Symposium (pp.11321137).
Xu, Y., Yang, J. Y., & Jin, Z. (2003). Theory analysis on FSLDA and ULDA. Pat-
tern Recognition, 36(12), 3031-3033.
Xu, Y., Yang, J. Y., Lu, J. F., & Yu, D. J. (2004). An effcient renovation on kernel
Fisher discriminant analysis and face recognition experiments. Pattern Recogni-
tion, 37(10), 2091-2094.
Yambor, W., Draper, B., & Beveridge, R. (2002). Empirical evaluation methods in
computer vision. Singapore World Scientifc Press.
Zhang, D., Kong, W. K., You, J., & Wong, M. (2003). Online palmprint identifca-
tion. IEEE Trans. Pattern Anal. Mach. Intell., 25(9), 1041-1050.
Zhou, X., & Bhanu, B. (2008). Feature fusion of side face and gait for video-based
human identifcation. Pattern Recognition, 41, 778795.
Matching Score Level Fusion 305
is prohibited.
Chapter XIV
Matching Score Level Fusion
Abstr Act
With this chapter we aims at describing several basic aspects of matching score
level fusion. Section 14.1 provides a description of basic characteristics of matching
score fusion in the form of introduction. Section 14.2 shows a number of matching
score fusion rules. Section 14.3 surveys several typical normalization procedures
of raw matching scores. Section 14.4 gives an example of matching score level
fusion method. Finally, Section 14.5 provides several brief comments on matching
score fusion.
14.1 Introduct Ion
Matching score level fusion is the most commonly used biometric information
fusion strategy because matching scores are easily available and because they retain
suffcient information to distinguish genuine matching from impostor matching.
Generally, a multi-biometric system based on the matching score level fusion works
as follows: each subsystem of the multi-biometric system exploits one biometric
trait to produce a matching score. Then these matching scores are normalized
and integrated to obtain the fnal matching score or fnal decision for personal
authentication. Studies on the basis and potential of score level fusion are very
helpful for us to understand and implement fusion at the score level. For example,
based on the likelihood ratio test Nandakumar, Chen, Dass and Jain (2008) proposed
a framework for the optimal combination of matching scores. Their study showed
that when fnite Gaussian mixture model was used to model the distributions of
is prohibited.
genuine and impostor match scores, the matching score level fusion could produce
good performance. Toh Kim, and Lee (2008b) studied the issue of optimizing the
ROC performance of the multimodal biometric system using the matching score
fusion strategy. Jain, Nandakumar, and Ross (2005) provided us with comprehensive
descriptions of various score normalization rules. Snelick, Uludag, Mink, Indovina
and Jain (2005) and Poh and Bengio (2006) presented the performances of matching
score level fusion algorithms and systems obtained using elaborate evaluation.
A number of studies (Dass, Nandakumar, & Jain, 2005; Kittler, Hatef, Duin, &
Matas, 1998; Poh & Bengio, 2006; Ross & Jain, 2003; Schmid, Ketkar, Singh, &
Cukic, 2006; Snelick Uludag, Mink, Indovina, & Jain, 2005; Vielhauer & Scheidat,
2005) have demonstrated that the matching score level fusion strategy can lead to
a higher accuracy than the single biometric system. A variety of cases of matching
score fusion have also been proposed (Brunelli & Falavigna, 1995; Doddington,
Liggett, Martin, Przybocki, & Reynolds, 1998; Islam, Mangayyagari, & Sankar,
2007; Kumar & Zhang, 2004; Scheidat, Vielhauer, & Dittmann, 2005; Tsalakanidou,
Malassiotis, & Strintzis, 2007; Tulyakov & Govindaraju, 2005; Yan & Bowyer,
2005). For example, Ribaric and Fratric (2005) acquired images containing both
fngerprints and palm prints and then used the extracted eigenpalm and eigenfnger
features to perform matching score level fusion. By conducting experiments on
a population approaching 1,000 individuals, Snelick, Uludag, Mink, Indovina
and Jain (2005) demonstrated that the multimodal fngerprint and face biometric
system, which combines the two biometric traits at the matching score level, was
signifcantly more accurate than any individual biometric systems.
In the context of verifcation, fusion at the matching score level can be carried out
using three distinct approaches. The frst approach treats fusion as a classifcation
problem (Ma, Cukic, & Singh, 2005) whereas the second approach treats fusion as
a combination problem. The third approach, namely the density-based approach,
treats fusion at the matching score level as a density-based score fusion problem,
which is based on the likelihood ratio test and explicit estimation of genuine and
impostor match score densities (Griffn, 2004). In the frst approach, the matching
score outputs of the individual biometric traits are frst concatenated to form a feature
vector. This feature vector is then classifed into one of two classes: genuine user
or impostor. This is also referred to as classifer-based score fusion (Nandakumar,
Chen, Dass and Jain, 2008). Examples of matching score level fusion using the frst
approach include fusion of face and speech data (Yacoub, Abdeljaoued, & Mayoraz,
1999), fusion of face and iris recognition (Wang, Tan, & Jain, 2003), fusion of face
and voice (Verlinde & Cholet, 1999), fusion of visual facial and acoustic features
(Chatzis, Bors, and Pitas, 1999), and fngerprint authentication (Jain, Prabhakar,
& Chen, 1999; Jain, Hong, & Bolle, 1997), etc. Note that the frst approach suffers
from the following issues: The frst issue is the unbalanced training set. That is,
is prohibited.
genuine match scores available for training are much fewer than available impostor
scores, since the numbers of available genuine match scores and impostor scores
are O(n) and O(n)
2
, respectively, where n is the number of users. The second issue
is that selecting and training a classifer that gives the optimal performance may be
diffcult. In the second approach, the individual matching scores are combined to
generate the fnal matching score which is then used to make the fnal decision. In
the following of this chapter, if there is no clear specifcation, the term matching
score fusion refers to fusion using the second approach mentioned above. The
density-based approach has the following advantage: If the score densities can be
estimated accurately, it is able to achieve optimal performance at any desired FAR.
A comparison of eight biometric fusion techniques conducted by NIST showed that
Product of Likelihood Ratios was most accurate (Ulery, Hicklin, Watson, Fellner,
& Hallinan, 2006). However, it also appeared that Product of Likelihood Ratios
was most complex to implement.
Generally a normalization step is used to normalize the matching scores gener-
ated from different biometric traits and then the normalized scores are integrated to
obtain the fnal matching score. One reason that the normalization step is adopted
is the matching scores of multiple biometric traits may not be homogeneous. For
example, it is possible that while a biometric trait yields a distance measure, an-
other biometric trait produces a similarity measure. In this case, it is obviously not
suitable to directly integrate two matching scores generated from the two biometric
traits. Another reason for using the normalization step is that the matching scores of
different biometric traits may have different numerical scales (ranges). For instance,
if the frst biometric trait yields scores in the range [10, 100] and the score from the
second biometric trait is in the range [0,1], directly fusing the scores without any
normalization will dramatically eliminate the contribution of the second biometric
trait. The third reason is that the matching scores produced by different subsystems
of a multi-biometric system may follow different statistical distributions. Due to these
reasons, a normalization procedure is essential to transform the matching scores of
individual biometric traits into a common domain prior to combining them.
Generally, a normalization procedure applied to raw matching scores has the
following two purposes: to convert the raw matching scores into ones that are of
the same kind of measurement, such as similarity or dissimilarity, and to make the
obtained matching scores of different biometric traits have a common range. A
good normalization procedure should be robust and effective. Robustness means
insensitivity to the presence of outliers. Effectiveness refers to the proximity of
the obtained estimate to the optimal estimate when the distribution of the data is
known. A variety of normalization procedures have been proposed for matching
score level fusion (Snelick, Uludag, Mink, Indovina, & Jain, 2005). The normalized
is prohibited.
matching scores will be further integrated to produce the fnal matching score for
personal authentication by using different approaches.
A number of fusion rules have been proposed for matching score level fusion. It is
likely that Simple-Sum, Min-Score and Max-Score are the most common matching
score level fusion rules. These rules appear to be not only simple but also effective.
In addition, Kittler, Hatef, Duin and Matas (1998) have tested several fusion rules
for the combination of face and voice, including sum, product, minimum, median,
and maximum rules. They also showed that the sum rule was not signifcantly
affected by the probability estimation errors. Dass, Nandakumar and Jain (2005)
combined the matching scores of multi-biometric traits based on generalized density
estimation. Jain and Ross (2002) and Brunelli and Falavigna (1995) respectively
proposed their user weighted fusion approaches, which assign different weights to
individual biometric traits. Lucey, Chen, Sridharan, and Chandran (2005), Chibelushi,
Deravi, and Mason (2002) and Dupont and Luettin (2000) exploited an exponential
weighting on the acoustic and visual a posteriori probability estimates to perform
audio-visual speech and speaker recognition. Toh, Kim, and Lee (2008a) conducted
biometric score fusion based on total error rate minimization. M. Montague and J.
A. Aslam (2001) proposed the relevance score normalization scheme. Fernandez,
Veldhuis, and Bazen (2006), Nandakumar, Chen, Jain, and Dass (2006) and Fier-
rez-Aguilar, Ortega-Garcia, Gonzalez-Rodriguez, and Bigun (2005) proposed to
design score normalization schemes that exploit quality information.
As presented above, matching score fusion has two key points, the normaliza-
tion procedure and the fusion approach using a certain fusion rule. In general, the
main steps of matching score fusion can be described as follows:
Step1. Raw matching scores of different biometric traits are converted into
the same kind of measurement.
Step2. The matching scores obtained using Step 1 are normalized by using
the normalization procedure.
Step3. The normalized matching scores are integrated to obtain the fnal
matching score or decision by using the fusion rule.
Step4. Personal authentication is performed based on the fnal matching score
or decision.
14.2 mAtch Ing score f us Ion r ules
Suppose that R biometric traits can be used for personal authentication and each
biometric trait produces a matching score. If these matching scores have been
normalized by the normalization procedure, they can be integrated. For the kth user,
is prohibited.
let
1 2
, ,...,
k k kR
S S S denote the normalized matching scores of his (her) R biometric
traits. The fnal matching score of the kth user can be obtained using the rules
shown in Sections 14.2.1 and 14.2.2.
14.2.1 general matching score f usion r ules
As presented in Section 14.1, matching score fusion rules integrate normalized
matching scores of a user to produce the fnal matching score. In this subsection
we present some widely used matching score fusion rules. These rules integrate
multiple normalized matching scores to produce the fnal scalar matching score.
Simple-Sumrule: TheSimple-Sum rule takes the sum of the R matching scores of
the kth user as the fnal matching score S
k
of this user. S
k
is calculated as follows:
1
R
k ki
i
S S
=
=
. (14.1)
Product rule: The Product rule regards the multiplication result of the R match-
ing scores of the kth user as the fnal matching score of this user, which is shown
as follows:
1,2,...,
k ki
i R
S S
=
=

(14.2)
In the case where the matching score means the a postpriori probability, the
product rule is essentially subject to the hypothesis that different biometric traits
are independent of each other, which ensures that the combinatorial postpriori
probability of multiple biometric traits equals the product of their respective
postpriori probabilities.
Min-Score rule: The Min-Score rule selects the minimum score from the R
matching scores of the kth user as the fnal matching score of this user. This rule
is expressed as follows:
1 2
min( , ,..., )
k k k kR
S S S S = . (14.3)
Max-Score rule: The Max-Score rule selects the maximum score from the
Rmatching scores of the kth user as the fnal matching score of this user. This rule
is shown as follows:
1 2
max( , ,..., )
k k k kR
S S S S = . (14.4)
is prohibited.
WeightedSumrule: The Weighted Sum rule assumes that the R biometric traits
have different signifcance in personal authentication and assigns different weights
to the matching scores of different traits. The weighted sum of the R matching
scores, which is shown in (14.5), is considered as the fnal matching score of the
kth user.
1
R
k n kn
n
S w S
=
=
, (14.5)
where w
n
represents the weight of the matching score of the nth biometric trait of
the kth user.
WeightedProductrule : Let w
n
stand for the weight of the matching score of
the nth biometric trait of the kth user. A Weighted Product rule can determine the
fnal matching score of the kth user using
1,...,
n
w
k kn
n R
S S
=
=

, (14.6)
The rules presented in this section can integrate multiple matching scores
representing one of the following measurements: a similarity measurement, the
a postpriori probability of the sample belonging to a certain class, or a distance
measurement between the sample and the registered template. The meaning of
the fnal matching score in personal verifcation can be described as follows. If
the integrated matching scores and the obtained fnal score are all probability or
similarity measurements, then the higher the fnal score between the sample and
the registered template of the claimed user is, the more likely the tested user is
the genuine subject. On the other hand, if the integrated matching scores and the
obtained fnal score are all distance measurements, then the smaller the fnal score
between the sample and the registered template of the claimed user is, the more
likely the tested user is the genuine subject. It should be pointed out that when the
raw matching scores denote different measurements, they cannot be integrated
directly. Instead, they should be frst converted into the same kind of measurement
and then be integrated.
14.2.2 Probability-Based Matching Score Fusion
Probability-based matching score fusion is also popular in the feld of biometrics.
We now will present several rules of probability-based matching score fusion to fuse
multi-biometric data at the matching score level from the viewpoint of classifcation.
In the context below we can fnd that these rules all integrate multiple biometric
traits by integrating their posteriori probabilities. Note that these rules such as the
probability-based Product, probability-based Max and probability-based Min rules
is prohibited.
are formally different from the general Product, Max-score and Min-score rules for
fusing matching scores as presented in Section 14.2.1. Indeed, probability-based
matching score fusion can be viewed as a special kind of fusion, which integrates
multiple matching scores to directly obtain the fnal authentication decision rather
than the fnal matching score. Probability-based matching score fusion is able to
do so because what it integrates are probabilities rather than ordinary matching
scores. Though probability-based matching score fusion directly produces the fnal
authentication decision and is similar to decision level fusion at this point, we do
not regard it as decision level fusion. The reason is that decision level fusion usually
combines multiple decisions to obtain the fnal decision whereas what probability-
based matching score fusion integrates are probabilities, a special kind of matching
score, rather than decisions.
Suppose that the biometric identifcation problem has m possible classes
1 2
, ,...,
m
and the multi-biometric system consists of R subsystems each exploiting
one biometric trait. Let x
i
, i = 1,2,...,R be the vector representing the ith biometric
trait. If ( )
k
p and ( | )
i k
p x respectively denote the a prior probability and the
conditional probability of x
i
given
k
, we can describe probability-based matching
score fusion rules as follows.
14.2.2.1 Probability-Based Product Rule
The following rule to combine probabilities of multiple biometric traits is called the
probability-based product rule (Kittler, Hatef, Duin, & Matas, 1998):
Assign
j
z , if
( 1) ( 1)
( ) ( | ) max ( ) ( | )
R R
j j i k k i
k
i i
p p x p p x
- - - -
=

. (14.7)
For a given sample represented by x
1
, x
2
,...,x
R
, this rule computes the product
( 1)
( ) ( | )
R
k k i
i
p p x
- -

with respect to each class
k
and then classifes the sample into the class that has
the maximum product value. Note that though both the probability-based Product
rule and the general Product rule as defned in (14.2) produce the fnal score based
on the idea that the fnal score and the normalized matching scores has a relation
of product, they are formally different. This is because the probability-based
Product rule is established on the basis of a more specifc knowledge background
as presented in the end of this section. The underlying rationale of (14.7) will be
also briefy shown in the end of this section.
is prohibited.
14.2.2.2 Probability-Based Sum Rule
The probability-based Sum rule combines probabilities of multiple biometric traits
to carry out classifcation for a sample denoted by x
1
, x
2
,...,x
R
as follows:
1 1
(1 ) ( ) ( | ) max((1 ) ( ) ( | ))
R R
j j i k k i
i i
k
R p p x R p p x
= =
- + = - +

.
(14.8)
Indeed, the probability-based Sum rule frst adds the sum of the a postpriori
probabilities with respect to each class of every biometric trait of a user to the
corresponding prior probabilities multiplied by (1 R). Then this rule classifes the
user into the class that has the maximum adding result.
14.2.2.3 Probability-Based Max Rule
The probability-based Max rule can be presented as follows:
Assign
j
z , if
1,..., 1,..., 1,...,
(1 ) ( ) max ( | ) max [(1 ) ( ) max ( | )]
j j i k k i
i R k m i R
R p R p x R p R p x
= = =
- + = - + .
(14.9)
14.2.2.4 Probability-Based Min Rule
The probability-based Min rule combines probabilities of multiple biometric traits
to perform classifcation as allows:
Assign
j
z , if
( 1) ( 1)
1,..., 1,..., 1,...,
( ) min ( | ) max ( ( ) min ( | ))
R R
j j i k k i
i R i R k m
p p x p p x
- - - -
= = =
= . (14.10)
14.2.2.5 Probability-Based Median Rule
According to the probability-based Median rule, probabilities of multiple biometric
traits should be combined for classifcation in the following way:
Assign
j
z , if
is prohibited.
1,...,
1,2,... 1,2,...
( ( | )) max ( ( | ))
j i k i
k m
i R i R
median p x median p x
=
= =
=
, (14.11)
where median denotes the median of the a posteriori probabilities.
Note that it has been demonstrated that the above probability-based rules to
combine different biometric traits all can be explained from the point of view of
Bayesian statistics (Kittler, Hatef, Duin, & Matas, 1998). For example, the condition
that the representations x
1
, x
2
,...,x
R
for different biometric traits of a sample are con-
ditionally independent, can allow the probability-based Product rule to be produced.
Another example is that under the condition that the a posteriori probabilities of
x
1
, x
2
,...,x
R
do not deviate dramatically from the prior probabilities, we are easy to
obtain the probability-based Sum rule.
We now will briefy show that how the probability-based Product rule can be
generated from the Bayesian decision theory. Indeed, if we apply the Bayesian
decision theory to the pattern recognition problem, we will classify a pattern z
described by vectors x
1
,x
2
,, x
R
using the following decision:
Assign
j
z if
1 2 1 2
( | , ,..., ) max ( | , ,..., )
j R k R
k
p x x x p x x x = . (14.12)
1 2
( | , ,..., )
k R
p x x x can be evaluated using
1 2
1 2
1 2
( , ,..., | ) ( )
( | , ,..., )
( , ,..., )
R k k
k R
R
p x x x p
p x x x
p x x x
=
. (14.13)
(14.13) also shows that for given x
1
, x
2
,...,x
R
,
1 2
( , ,..., )
R
p x x x has the same effect
on the value of the
1 2
( | , ,..., )
k R
p x x x . Therefore, rule (14.12) is equivalent to the
following rule:
Assign
j
z if
1 2 1 2
( , ,..., | ) ( ) max ( , ,..., | ) ( )
R j j R k k
k
p x x x p p x x x p = . 14.14)
Suppose that the representations x
1
, x
2
,...,x
R
are conditionally independent, then
we get
1 2
( , ,..., | ) ( | )
R k i k
i
p x x x p x =
, (14.15)
is prohibited.
Combining (14.14) and (14.15), we obtain the following rule:
Assign
j
z if
( ) ( | ) max ( ) ( | )
j i j k i k
k
i i
p p x p p x =

. (14.16)
We also know that ( | ) ( ) ( | ) ( )
i k k k i i
p x p p x p x = . This allows rule (14.16) to
become
Assign
j
z if
( 1) ( 1)
( ( )) ( ) ( | ) max( ( )) ( ) ( | )
R R
i j j i i k k i
k
i i i i
p x p p x p x p p x
- - - -
=

.
(14.17)
Because (14.17) shows that for given x
1
, x
2
,...,x
R
,
( )
i
i
p x

has no infuence on the classifcation decision, we can eliminate it from this rule
and obtain (14.7).
As presented early, probability-based matching score fusion is a special kind of
matching score fusion. Indeed, we can consider that probability-based matching
score fusion implicitly implements the following two procedures: the frst procedure
of integrating probabilities of multiple biometric traits and the second procedure
of using the integrated probabilities and the Bayesian decision theory for fnal
decision-making. This, therefore, makes probability-based matching score fusion
formally different from ordinary matching score fusion, which usually integrates
matching scores of multiple biometric traits to produce the fnal matching score
and no authentication decision will be made.
14.3 norm Al IzAt Ion Procedures of mAtch Ing
scores
As presented in Section 14.1, normalization of matching scores is very critical for
matching score fusion. This subsection focuses on some popular score normalization
procedures.
Min-Max normalization procedure, which is best suited for the case where the
bounds of the matching score of each biometric trait is known, is one of the simplest
is prohibited.
score normalization procedures. This procedure maps raw scores to the [0, 1] range
using (14.18). The quantities max(S) and min(S) denote the upper and lower bounds
of the score range, respectively.
'
min( )
max( ) min( )
k
k
S S
S
S S
-
=
-
. (14.18)
Z-scorenormalization procedure aims at transforming the raw matching score
to a distribution having mean of 0 and standard deviation of 1. The transformation
is defned as follows:
'
( )
( )
k
k
S mean S
S
std S
-
=
, (14.19)
where mean(S) and std(S) are respectively used to denote the arithmetic mean
and standard deviation operators. It should be pointed out that because both the
mean and standard deviation are sensitive to outliers, this procedure is not robust.
As a result, the Z-score normalization procedure does not guarantee that the nor-
malized scores have a common numerical range.
Tanhnormalization procedure maps the raw matching scores to the (0, 1) range
using the following equation:
'
1 ( )
[tanh(0.01 ) 1]
2 ( )
k
S mean S
S
std S
-
= + , (14.20)
where mean(S) and std(S) also denote the arithmetic mean and standard devia-
tion operators, respectively.
Medianandmedianabsolutedeviation normalization procedure normalizes
raw matching scores as follows:
'
( )
(| ( ) |)
k
k
k
S median S
S
median S median S
-
=
-
, (14.21)
where median(S) stands for the median value of S. Indeed, the term median means
the middle of a distribution. That is, half the scores are above the median and the
is prohibited.
other half are below the median. The median is less sensitive to extreme scores
than the arithmetic mean and this makes it be a better measure. This procedure is
insensitive to outliers and it is therefore more robust than the Z-score normaliza-
tion procedure.
Before we discuss the next normalization procedure, we analyze the error of
personal authentication using a single biometric trait. The errors caused by individual
biometric traits stem from the overlap of the genuine and impostor score distribu-
tions. This overlap region can be partially described by its center c and its width
w. The next normalization procedure (logistic normalization) is able to decrease
the effect of the overlap and to increase the separation of the genuine and impostor
distributions, while it still maps the scores to the [0,1] range.
Logistic normalization procedure uses a logistic function to normalize raw
matching scores. This logistic function is defned as
'
1
1
mm
k BS
S
A e
-
=
+
, (14.22)
where S
nm
denotes the normalized matching score obtained using the Min-Max
normalization procedure. Note that the constants A and B are respectively calcu-
lated using
1
1 A = -
and
ln A
B
c
= .
Figure 14.1. Illustration ofthe logistic normalization procedure
is prohibited.
Figure 14.2. Illustration of the Quadric-Line-Quadric procedure

w
c (1,0) (0,0)
(0,1)
MM
n
AD
n
The constant is selected to be a small value. This specifcation results in the infec-
tion point of the logistic function occurring at c, i.e. the center of the overlap region.
A logistic normalization function defned as in (14.22) is shown in Figure 14.1.
Quadric-Line-Quadricprocedureleavesthe overlap zone, with center c and
width w, of the genuine and impostor score distributions unchanged and transforms
the other regions using two quadratic functions. This procedure is formulated as
(14.23) and illustrated as Figure 14.2.
2
'
1
, ( 2)
2
, ( 2) ( 2)
( 2) (1 2)( 2),
nm nm
k nm nm
nm
S S c w
c w
S S c w S c w
c w c w S c w otherwise
= - < < +
+ + - - - -
(14.23)
where S
nm
still denotes the normalized matching score obtained using the Min-Max
normalization procedure.
Jain, Nandakumar and Ross (2005) studied the performance of different normal-
ization techniques and fusion rules in the context of a multimodal biometric system
based on the face, fngerprint and hand-geometry traits of a user. Their experiments
on a database of 100 users indicate that the applications of the minmax, z-score,
is prohibited.
and tanh normalization rules followed by a simple sum scheme of score fusion
obtain better recognition performance than other methods. However, experiments
also show that the minmax and z-score normalization techniques were sensitive
to outliers in the data.
14.4 exemPl If IcAt Ion: Inform At Ion f us Ion of f Ace
And PAlm Pr Int
This section presents a scheme for fusing face and palm print information to conduct
personal authentication. The used information fusion strategy is the frst approach
of the matching score level fusion, which considers the matching score outputs of
the individual biometric traits as a feature vector and then personal authentication
is performed on the basis of this feature vector. The proposed scheme also uses the
claimed identity of users as a feature for fusion. The experimental results show that
the facial and palm print images can be simultaneously acquired by using a pair of
digital cameras and can be integrated to obtain a high accuracy of personal authen-
tication. This section also analyzes and evaluates the performance of a personal
authentication system using two-class separation criterion functions. The scheme
proposed in this section may also be applicable to other multi-biometric systems.
14.4.1 introduction
This section investigates a bimodal biometric system using face and palm print.
Face has highest user acceptance and its acquisition is most convenient to users
(Prabhakar, Pankanti, & Jain, 2003). Face and palm print images can be conveniently
acquired from a digital camera. One of the important features that is only available
in personal authentication, but not in recognition, is the claimed user identity. The
claimed user identity is unique for every user and can be used to restrict the deci-
sion space, i.e. range of matching scores, in user authentication. The claimed user
identity can be suitably coded and then used as a feature to classify the genuine
and impostor matching scores and is investigated in this section.
14.4.2 Proposed System and Method
The block diagram of the proposed bimodal biometric authentication system is
shown in Figure 14.3. The acquired grey-level images from the palm print and face
are presented to the system. The matching scores generated from each of the two
biometric traits are used as inputs of a neural network classifer. Because each user
has its claimed identity, the claimed user identity is also used as a feature to neural
is prohibited.
network classifer. The trained neural network classifes the user as a genuine user
or impostor.
14.4.2.1 Computation of Face Matching Scores
We use the eigenface (Turk & Pentland, 1991) algorithm, a popular appearance
based face representation algorithm to transform face images into lower-dimen-
sional data. Original face images acquired from all the subjects are of size M N
and each image can be represented by a one-dimensional vector of MN dimension
using row ordering. Principal component analysis (PCA) is applied to the normal-
ized set of such training vectors to produce a set of orthogonal vectors, also known
as eigenfaces. Note that the projection is in the form of a vector. The projection of
a face image on eigenfaces is used as features of the face image.
The matching score for features (x
q
) of each test face image is obtained by cal-
culating the similarity between the features (x
q
) of the claimed identity and features
(x
c
) of a training sample as follows.
Figure 14.3. Personal authentication using face and palm print
is prohibited.
1
|| || . || ||
T
q c
q c
x x
x x
=
, (14.24)
14.4.2.2 Computation of Palm Print Matching Scores
Palm print contains several complex features, e.g. minutiae, principal lines, wrinkles
and texture. The palm print matching approach used in this work is same as de-
tailed in (Kumar, Wong, Shen, & Jain, 2003). Four directional spatial masks are
used to capture line features from each of the palm print images. The combined
directional map is generated from voting of the resultant four images. The standard
deviation of pixels, from each of the 24 24 pixel overlapping block with 25%
overlap, in the combined image is used to form the feature vector. The palm print
matching scores are generated by computing the similarity measure
2
, similar to
(14.24), between the feature vectors from a testing image and those stored during
the training phase.
14.4.2.3 Neural Networks-Based Classifcation
The matching scores from the face and palm print are used as inputs to train a
feed-forward neural (FFN) network. We also use the claimed identity of every user
as an input of the FFN. A three-layer FFN with P
l
neurons in the lth (l = 1,2,...,Q)
layer is based on the following architecture (Kumar, 2003):
1 1, 1 1 1
1
, ( )
l
P
l l l l l l
j ij i i i
i
w y y g
- - - - -
=
= =
, (14.25)
where the sum of weighted inputs for the jth( j = 1,2,...,P
l
) neuron in the lth(l =
1,2,...,Q) layer is represented by
l
j
. The weights from the ith neuron at the (l = 1)th
layer to the jth neuron in the lth layer are denoted by
1, l l
ij
w
-
and
l
j
y
is the output for
the jth neuron in the lth layer. The values 1 and 1, corresponding to impostor and
genuine responses, were given to the FFN during training as the correct output
responses for expected classifcation during the training. As defned as in (14.26),
the hyperbolic tangent sigmoid activation function is empirically selected for frst
two layers, while a linear activation function is chosen for third layer.
3 3
( ) tanh( ), 1, 2
( ) ( )
l l
j j
j j
g l
g a
= =
. (14.26)
is prohibited.
The back-propagation training algorithm is used for minimizing the target
function T
e
defned by:
2
1 1
1
( )
Q
K P
Q
e jk jk
k j
Q
T y o
KP
= =
= -

, (14.27)
where k is an index for input-output pair and
2
( )
Q
jk jk
y o - is the squared difference
between the actual output value at the jth output layer neuron for pair k and the
target output value. The FFN uses Levenberg-Marquardt algorithm (Masters, 1995)
and a constant learning rate to update the connection weights
1, l l
ij
w
-
.
14.4.3 Experiment and Results
The proposed method was investigated on the ORL face database (please refer to
Section 3.2.3) from 40 subjects with 10 images per subject. The hand images from 40
subjects, with 10 images per subject, were acquired by using a digital camera. Each
of the subjects for palm print and face were randomly paired to obtain a bimodal
set for every subject. For the palm print image, a 300 300 region of interest was
automatically segmented and the 144-dimensional feature vector was extracted as
detailed in Kumar, Wong, Shen and Jain (2003). The eigenface algorithm was ap-
plied to produce 40-dimensional feature vector for every face image. The matching
scores for face and palm print were computed by using similarity measure (14.24).
The frst four images samples, from face and palm print, were used for training
and rest six were for testing. Then genuine and impostor matching scores from the
training samples were used to train the 18/5/1 FFN. The learning rate was fxed at
0.01. Because a neural network cannot guarantee that the obtained training error
is global, the FFN was trained 10 times with the same parameters and the result
with the smallest of training errors of all the results are reported. The trained
neural network was used to test 240 (406) genuine and 9360 (40396) impos-
tor matching scores from the test data. The claimed user identity was also used
as an input component of FFN. The experimental results are shown in Table 14.1.
From Table 14.1, we know that the total error is 1.5% (=0.7%+0.8%). However, if
the claimed user identity was not exploited as an input component of the FFN, the
total minimum error obtained by the FFN would be 2.8 %. In order to ascertain the
improvement (or degradation) in the separation of the FFN outputs on the genuine
and impostor, the performance indices using three objective functions (Kumar &
Pang, 2002), were considered.
is prohibited.
1
g
i
J =
,
2
2
( )
g i
g i
J
-
=
,
2
3
2 2
( )
g i
g i
J
-
=
+
(14.28)
where
g
and
i
represent the means of the FFN outputs on the genuine and imposter
class, respectively, and
g
,
i
are the standard deviations of the FFN outputs on the
genuine and imposter class. Note that J
1
and J
2
do not account for the variance of
the FFN outputs whereas J
3
simultaneously account for the mean and variance of
the FFN outputs. These indices were computed from the test data and are displayed
in Table 14.2. The bracketed entries in this table show the respective indices when
the claimed identity of user is not used as an input component of the FFN. These
entries can be used to interpret the performance increase when the claimed user
identity is used as an input component of the FFN. Table 14.2 also shows the equal
error rate (EER) for each of the corresponding cases.
14.4.4 conclusions
Palm print and face images can be simultaneously acquired and used to perform
personal authentication. The experimental result shows that the proposed bimodal
Table 14.1. Experimental results of personal authentication using face and palm
print
FAR FRR
Face 3.0 10
Palm print 3.8 3.2
Experimental results using face and palm
print
0.7 0.8
Table 14.2. Performance indices of personal authentication using face and palm
print
J
1
J
2
J
3
EER
Face 3.85(1.05) 2.11(0.00) 4.42(2.34) 8.33%(8.69%)
Palm print 4.38(1.03) 2.61(0.00) 8.61(3.71) 3.65%(4.32%)
Experiment using face
and palm print
4.84(4.78) 3.04(2.99) 35.57(23.78) 0.84%(2.09%)
is prohibited.
system can achieve a higher accuracy than the single biometric system using palm
print or face images. J
3
appears to be an effective performance index for evaluating
the improvement in performance of the multi-biometric system. The experimental
result also shows that the claimed user identity has signifcant effect in improving
performance of the biometric system. This improvement is attributed to the fact
that the FFN classifer uses the claimed user identity to reduce the uncertain range
of the output value for the corresponding user.
14.5 comments
Fusion at the matching score level is the most popular fusion strategy in the feld
of multi-biometrics. The performance of this fusion strategy can be seen in numer-
ous studies. The application presented in Section 14.3 uses the second approach of
matching score fusion. Note that besides different biometric traits can be fused at
the matching score level; matching score level fusion may be also performed for
two distinct kinds of features generated from the same biometric trait such as the
shape and texture features of tongue-print (Zhang, Liu, Yan, & Shi, 2007). There
are some preferred normalization procedures and fusion rules. For example, a
variety of applications has shown the satisfactory performance of the simple sum
rule in increasing the system accuracy. Studies on Min-max and z-score, two
extensively applied normalization procedures, showed that they could work well
in a large portion of real-world cases. It is also known that some normalization
schemes will work well if the scores follow a specifc distribution. For example,
z-score normalization is optimal if the scores of all the biometric traits follow the
Gaussian distribution. However, no matching score fusion rule or normalization
procedure can work well under all circumstances. In this sense, it is worthwhile
to explore the possibility of determining and applying different normalization
techniques or score fusion rules to matching scores of different biometric traits. It
is also considered that setting different weights for the matching scores of differ-
ent biometric traits (Griffn, 2004; Kumar, Wong, , Shen, & Jain, 2003) allows the
multi-biometric system to perform better than the systems that assign the same set
of weights to all the biometric traits.
r eferences
Brunelli, R., & Falavigna, D. (1995). Person identifcation using multiple cues. IEEE
Trans. Patt. Anal. Machine Intell., 12(10), 955-966.
is prohibited.
person authentication. IEEE Trans. Systems, Man, and Cybernetics, Part A: Systems
and Humans , 29(6), 674-681.
Chibelushi, C. C., Deravi, F., & Mason, J. S. D. (2002). A review of speech-based
bimodal recognition. IEEE Trans. Multimedia, 4(1), 23-37.
Dass, S. C., Nandakumar, K., & Jain, A. K. (2005). A principled approach to score
level fusion in multimodal biometric systems. Proc. of Audio- and Video-based
Biometric Person Authentication (AVBPA) (pp.1049-1058).
Doddington, G., Liggett, W., Martin, A., Przybocki, M., & Reynolds,D. (1998).
Sheeps, goats, lambs and wolves: A statistical analysis of speaker performance in
the NIST 1998 speaker recognition evaluation. Proceedings of the 5th International
Conference on Spoken Language Processing, Sydney, Australia.
Dupont, S., & Luettin, J. (2000). Audio-visual speech modeling for continuous
speech recognition. IEEE Trans. Multimedia, 2(3), 141-151.
Fernandez, F. A., Veldhuis, R., & Bazen, A. M. (2006). On the relation between
biometric quality and user-dependent score distributions in fngerprint verifcation.
Proc. of the Second International Workshop on Multimodal User Authentication,
Toulouse, France, May 2006.
Fierrez-Aguilar, J., Ortega-Garcia, J., Gonzalez-Rodriguez, J., & Bigun, J. (2005).
Discriminative multimodal biometric authentication based on quality measures.
Griffn, P. (2004). Optimal biometric fusion for identity verifcation. ( Identix Cor-
porate Research Center, Tech. Rep. RDNJ-03). New Jersey.
Islam, T., Mangayyagari, S., & Sankar, R. (2007). Enhanced speaker recognition
based on score level fusion of AHS and HMM. Proceedings of IEEE SoutheastCon.
(pp. 1419).
Jain, A. K., Hong, L., & Bolle, R. (1997). On-line fngerprint verifcation. IEEE
Trans. Patt. Anal. Machine Intell., 19(4), 302-314.
Jain, A. K., Nandakumar, K., & Ross, A. (2005). Score normalization in multimodal
biometric systems. Pattern Recognition, 38(12), 2270-2285.
Jain, A. K., Prabhakar, S., & Chen, S. (1999). Combining multiple matchers for a
high security fngerprint verifcation system. Pattern Recognition Letters, 20(11-
13), 1371-1379.
is prohibited.
Jain, A. K., & Ross, A. (2002). Learning use-specifc parameters in a multibiometric
system. Proc. of International Conference on Image Processing (ICIP), (pp. 22-
25).
Kittler, J., Hatef, M., Duin, R. P. W., & Matas, J. (1998). On combining classifers.
IEEE Trans. Patt. Anal. Machine Intell., 20(3), 226-239.
Kumar, A. (2003). Neural network based detection of local textile defects. Pattern
Recognition, 36(7), 1645-1659.
Kumar, A., & Pang, G. (2002). Defect detection in textured materials using optimized
flters. IEEE Trans. Systems, Man, and Cybernetics: Part B, Cybernetics, 32(5),
553-570.
Kumar, A., Wong, D. C. M., Shen, H., & Jain, A. K. (2003). Personal verifca-
tion using palmprint and hand geometry biometric. Proceedings of Audioand
Videobased Biometrie Person Authentication (pp. 668-678).
Kumar, A., & Zhang, D. (2004). Palmprint authentication using multiple classifers.
Proc. SPIE Symposium on Defense & Security - Biometric Technology for Human
Identifcation (pp. 20-29).
Lucey, S., Chen, T., Sridharan, S., & Chandran, V. (2005). Integration strategies
for audio-visual speech processing: applied to text-dependent speaker recognition.
IEEE Trans. Multimedia, 7(3), 495-506.
Ma, Y, Cukic, B., & Singh, H. (2005). A classifcation approach to multibiometric
score fusion. Proc. Fifth Intl Conf. Audio Video-Based Biometric Person Authen-
tication (pp. 484-493).
Masters, T. (1995). Advanced algorithms for neural networks: A C++ source book.
New York: John Wiley & Sons, Inc.
Montague, M., & Aslam, J. A. (2001). Relevance score normalization for metasearch.
Proc of Tenth International Conference on Information and Knowledge Manage-
ment (pp. 427-433).
Nandakumar, K., Chen, Y., Dass, S. C., & Jain, A. K. (2008). Likelihood ratio-based
biometric score fusion. IEEE Trans. Patt. Anal. Machine Intell., 30(2), 342-347.
Nandakumar, K., Chen, Y., Jain, A. K., & Dass, S. (2006). Quality-based score
level fusion in multibiometric systems. Proc. Intl Conf. Pattern Recognition (pp.
473-476).
is prohibited.
Poh, N., & Bengio, S. (2006). Database, protocol and tools for evaluating score-
level fusion algorithms in biometric authentication. Pattern Recognition, 39(2),
223-233.
Prabhakar, S., Pankanti, S., & Jain, A. K. (2003). Biometric recognition: security
and privacy concerns. IEEE Spectrum, 1(2), 33-42.
Ribaric, S., & Fratric, I. (2005). A biometric identifcation system based on eigenpalm
and eigenfnger features. IEEE Trans. Patt. Anal. Machine Intell., 27(11), 1698-
1709.
Ross, A., & Jain, A. K. (2003). Information fusion in biometrics. Pattern Recogni-
tion Letters, 24(13), 2115-2125.
Scheidat, T., Vielhauer, C., & Dittmann, J. (2005). Distance-level fusion strategies
for online signature verifcation. Proc. IEEE Int. Conf. on Multimedia and Expo
(ICME) (pp. 1294-1297).
Schmid, N. A., Ketkar, M., Singh, H., & Cukic, B. (2006). Performance analysis of
iris-Based identifcation system at the matching score level. IEEE Trans. Informa-
tion Forensics and Security, 1(2), 154-168.
Snelick, R., Uludag, U., Mink, A., Indovina, M., & Jain, A. (2005). Large-scale
evaluation of multimodal biometric authentication using state-of-the-art systems.
IEEE Trans. Patt. Anal. Machine Intell., 27(3), 450-455.
Tsalakanidou, F., Malassiotis, S., & Strintzis, A. (2007). A 3D face and hand bio-
metric system for robust user-friendly authentication. Pattern Recognition Letters,
28(16), 2238-2249.
Toh, K. A., Kim, J., & Lee, S. (2008a). Biometric scores fusion based on total error
rate minimization. Pattern Recognition , 41(3), 1066-1082.
Toh, K. A., Kim, J., & Lee, S. (2008b). Maximizing area under ROC curve for
biometric scores fusion. Pattern Recognition, 41(11), 3373-3392.
Tulyakov, S., & Govindaraju, V. (2005). Combining matching scores in identifca-
tion model. Proc. of Eighth International Conference on Document Analysis and
Recognition (pp. 1151-1155).
Turk, M. A., & Pentland, A. P., (1991). Eigenfaces for recognition. J. Cognitive
Neuroscience, 3(1), 71-76.
Ulery, B., Hicklin, A. R., Watson, C., Fellner, W., & Hallinan, P. (2006). Studies
of biometric fusion(NIST). Gaithersburg, MD: National Institute of Standards and
Technology.
is prohibited.
verifcation application. Proc. of Second International Conference on AVBPA (pp.
188-193).
Vielhauer, C., & Scheidat, T. (2005). Multimodal biometrics for voice and handwriting.
Proceedings of the 9th International Conference on Communications and Multi-
media Security (pp. 191-199).
identity verifcation. Proc. of Fourth International Conference on AVBPA (pp.
805-813).
Yacoub, S. B., Abdeljaoued, Y., & Mayoraz, E. (1999). Fusion of face and speech
data for identity verifcation. IEEE Trans. Neural Networks, 10(5), 1065-1075.
Yan, P., & Bowyer, K. W. (2005). MultiBiometrics 2D and 3D ear recognition.
Proceedings of Audioand Videobased Biometrie Person Authentication (pp.
503-512).
Zhang, D., Liu, Z., Yan, J., & Shi, P. (2007). Tongue-Print: A novel biometrics pattern.
Proceedings of the International conference on Biometrics (pp.1174-1183).
is prohibited.
Chapter XV
Decision Level Fusion
Abstr Act
With this chapter, we frst present a variety of decision level fusion rules and clas-
sifer selection approaches, and then show a case study of face recognition based
on decision level fusion, and fnally offer a summary of three levels of biometric
fusion technologies. In a multi-biometric system, classifer selection techniques may
be associated with the decision level fusion as follows: classifer selection is frst
carried out to select a number of classifers from all classifer candidates. Then the
selected classifers make their own decisions and the decision level fusion rule is
used to integrate the multiple decisions to produce the fnal decision. As a result, in
this chapter, we also introduce classifer selection by showing a classifer selection
approach based on correlation analysis. This chapter is organized as follows. Section
15.1 provides an introduction to decision level fusion. Section 15.2 presents several
simple and popular decision level fusion rules such as the AND, OR, RANDOM,
Voting rules, as well as the weighted majority decision rule. Section 15.3 introduces
a classifer selection approach based on correlations between classifers. Section
15.4 presents a case study of group decision-based face recognition. Finally, Section
15.5 offers some comments on three levels of biometric fusion.
15.1 Introduct Ion
Though the term decision level fusion has appeared widely in the biometric
literature, it is not used only in the feld of biometrics. Indeed, as an information
fusion strategy, decision level fusion also has been widely applied in a number of
Decision Level Fusion 329
is prohibited.
areas such as multisensor data fusion (Hall & Llinas, 1997), multispectral image
fusion and geoscience data fusion (Jeon & Landgrebe, 1999; Fauvel, Chanussot,
& Benediktsson, 2006). We would like to regard decision level fusion as a term
of information science rather than a term of biometrics. In some cases on multi-
biometrics, the term symbol level fusion (Tien, 2003; Gee & Abidi, 2000; Dop,
1999) is also used to represent decision level fusion. The decision level fusion
strategy integrates biometric information in a simple and straightforward way in
comparison with feature level fusion, which usually directly integrates different
biometric traits at the feature level, and matching score level fusion which usually
requires that before fusion the matching scores of different biometric subsystems
be normalized. A system using the decision level fusion strategy integrates differ-
ent biometric data at a later stage than the multi-biometric system using feature
level fusion or matching score level fusion strategies. The multi-biometric system
using the decision level fusion strategy can be described as follows: The system
consists of a number of biometric subsystems each of which uses a biometric trait
and makes the authentication decision independently. The decision level fusion
strategy is then used to combine the decisions of the biometric subsystems to pro-
duce the fnal decision.
Various decision level fusion methods such as Boolean conjunctions, weighted
decision methods, classical inference, Bayesian inference, and DempsterShafer
method (Jain, Lin, Pankanti, & Bolle, 1997), voting (Zuev & Ivanon, 1996) have
been proposed. Prabhakar and Jain (2002) combined classifer selection and deci-
sion level fusion techniques to perform fngerprint verifcation. Hong and Jain
(1998) integrated faces and fngerprints at the decision level. Chatzis, Bors, and
Pitas (1999) used fuzzy clustering algorithms to implement decision level fusion.
Osadciw, Varshney and Veeramachaneni (2003) proposed a Bayesian framework
to perform decision fusion based on multiple biometric sensors. In addition, modi-
fed KNN approach (Teoh, Samad, & Hussain, 2002), decision trees and logistic
regression (Verlinde & Cholet, 1999) were also used to fuse multiple biometric
traits at the decision level. More studies on decision level fusion such as fusion of
iris and face, fusion of 3D data can be found in (Wang, Tan, & Jain, 2003; Gkberk
& Akarun, 2006; Li, Zhao, Ao, & Lei, 2005; Gokberk, Salah, & Akarun, 2005;
Freedman, 1994; Teoh, Samad, & Hussain, 2004; Niu, Han, Yang, & Tan, 2007).
It should be noted that the theoretical framework described by Kittler, Hatef, Duin
and Matas (1998) is able to derive a number of real rules for combining classifers.
Roli, Kittler, Fumera, and Muntoni (2002) classifed the decision fusion strategies
into two main classes: fxed and trained rules. Fusion strategies such as majority
voting and the sum rule are recognized as fxed rules. These strategies might al-
low combination of different systems with similar performance to perform well.
Some techniques such as weighted averaging and behavior knowledge space are
is prohibited.
examples of trained rules, which may allow combination of systems with different
performance to improve authentication performance.
With this chapter, we mainly analyze and discuss a variety of decision fusion
rules, classifer selection approaches, and a case study of face recognition based on
decision fusion. In the following section, we will present some simple and popular
decision fusion rules such as the AND, OR, RANDOM, Voting rules. Also, we
will present weighted majority decision rule and classifcation combination based
on the confusion matrix with suffcient description.
15.2 r ules And methods of dec IsIon l evel f us Ion
Here we briefy introduce some typical rules and methods proposed for decision
level fusion.
AND rule: This rule allows a user to be accepted as a genuine user only if all
of the decisions of different subsystems using different biometric traits provide
positive authentication results. Usually the method using this rule produces a low
false acceptance rate whereas it also results in a high false rejection rate. This can
be shown as follows: suppose that a multi-biometric system using the AND rule
consists of two biometric subsystems each exploiting one of two single biometric
traits and the two biometric traits are independent of each other. If the false accept
rate and false reject rate of the frst subsystem are respectively FAR
1
and FRR
1
and
the two rates of the second subsystem are respectively FAR
2
and FRR
2
, we can
consider that the false accept and false reject rates of the multi-biometric system
are FAR
1
* FAR
2
and FRR
2
, respectively. Because FAR
1
, FRR
1
, FAR
2
and FRR
2
are
all smaller than 1, the multi-biometric system has a lower false accept rate than
both of the two subsystems, whereas the multi-biometric system has a higher false
reject rate than both of the two subsystems.
OR rule: This rule means that a user will be treated as a genuine user by a
multi-biometric system if either of his/her multiple biometric traits is accepted by
the system. Generally, this rule produces a lower false rejection rate and a higher
false acceptance than each of the biometric subsystems. For the case where there
are two independent biometric traits, we can consider that the false reject and false
accept rates of the multi-biometric system using the OR rule are FRR
1
* FRR
2
and
FAR
1
+ FAR
2
FAR
1
* FAR
2
, respectively. Here FAR
1
and FRR
1
are still the false
accept rate and false reject rate of the frst biometric subsystem, respectively. FAR
2

and FRR
2
are also the false accept rate and false reject rate of the second biometric
subsystem, respectively. This implies that the multi-biometric system has a lower
false reject rate than both of the two subsystems, whereas it has a higher false ac-
cept rate than both of the two subsystems.
is prohibited.
RANDOMrule:This rule allows a biometric trait to be randomly chosen to
perform personal authentication. This very simplistic idea can make it harder for
intruders to spoof the system. Moreever it comes without the inconvenience of a
multilevel data acquisition for each authentication attempt.
Voting: When voting techniques are used, the global decision rule is obtained
simply by fusing the decisions made by subsystems using different biometric traits.
Majority voting is a popular voting method. In this method, the class voted by most
of the subsystems will be regarded as the result of a fusion decision. If no class
won more than half of the votes, the input may be rejected. The method is simple
and easy to realize.
Rank Method: This kind of method is used to fuse information of multiple
biometric traits in the case where the classifers produce a ranked list of class
labels. The Borda count method is a widely used rank-based fusion scheme. This
method can be used for single-winner elections in which each voter rank-orders
all the candidates. The Borda count method calculates the combined ranking by
summing the class ranks as assigned by the individual voters. The candidate with
the maximum ranking score will be accepted as the winner. More specifcally, we
describe the Borda Count method for an election as follows: Each candidate gets
1 point for each last place vote received, 2 points for each next-to-last point vote,
etc., all the way up to N points for each frst place vote (where N is the number of
candidates/alternatives). The candidate with the largest point total wins the elec-
tion. If the Borda Count method is applied to a 4 candidate election, each 4th place
vote is worth 1 point, each 3rd place vote is worth 2 points, each 2nd place vote is
worth 3 points, and each 1st place vote is worth 4 points.
Note that in the Borda Count Method for multi-biometric verifcation, there are
only two candidates, i.e. genuine user and imposter. Therefore, there are only
two point numbers, 1 and 2. Each subsystem of a multi-biometric system actually
acts as a voter. A tested biometric trait is classifed into one candidate (genuine
user or imposter) that has the maximum sum of points.
15.2.6 Weighted Majority Decision Fusion Rule
15.2.6.1 Idea of Weighted Majority Decision Fusion
If there are p available biometric traits, decision level fusion for multi-biometric
can be viewed as an issue to make an optimum fnal decision u
0
under the condi-
tion that the decisions {u
1
,
...
,

u
p
} respectively for all single biometric traits are all
given. Note that u
1
is made by using only the frst biometric trait, u
2
is made by
using only the second biometric trait, etc., all the way up to u
p
. {u
1
,
...
,

u
p
} are called
local decisions. It should be pointed out that among u
1
,
...
,

u
p
some decisions may be
is prohibited.
identical. We can determine a decision fusion method based on a cost function.
The idea to design the cost function is as follows (Jeon & Landgrebe, 1999): among
the local decisions {u
1
,
...
,

u
p
}, some of the decisions are more dependable in terms
of reliability than the others. Therefore, it would be desirable if the fnal decision
u
0
is as consistent as possible with those reliable local decisions. To implement the
idea of emphasizing reliable local decisions among {u
1
,
...
,

u
p
}, we can defne the
following cost function:
0 1 0
1
( ; ,..., , ) ( ; , )
p
p j k j
k
J u u u J u u
=
=
, (15.1)
where
j
is the genuine class of the user of the multi-biometric system,
0
( ; , )
k j
J u u
is a cost function associated with the kth biometric traits. Indeed,
0
( ; , )
k j
J u u shows
the cost given to an action of selecting u
0
based on the local decision u
k
. Eq.(15.1)
regards the sum of all the local costs as the actual cost of the action of selecting u
0

based on {u
1
,
...
,

u
p
}.
We can determine the fnal decision based on a cost function
0
( ; , )
k j
J u u
sat-
isfying
0 0 0
( ; , ) ( ; ) 1 ( ) ( , )
k j k k k k
J u u J u u C u u u = = - , (15.2)
where 0 ( ) 1
k k
C u . The value of
0
( , )
k
u u is
0
0
1
( , )
0
k
k
u u
u u
othrwise
=
.
According to (15.2), we know that the cost of selecting the local decision u
k
as the
fnal decision u
0
is 1 C
k
(u
k
), whereas the cost in other cases is one. C
k
(u
k
), therefore,
can control relative importance of consistency between u
0
and u
k
. We can select
appropriate values for C
k
(u
k
) in such a way that the less reliable the decision u
k
is,
the less effect on making the fnal decision it has.
If we combine Eq.(15.1) and Eq.(15.2), we can know that
0 1 0
1
( ; ,..., , ) ( ) ( , )
p
p j k k k
k
J u u u p C u u u
=
= -
. Based on
0 1 1 0 1
{ ( ; ,..., , )} ( ,..., ) ( ; ,..., , ))
p j p p j
E J u u u P u u J u u u = , we can obtain the following
equation representing the expected cost (Jeon & Landgrebe, 1999):
0 1 1 0
1
{ ( ; ,..., , )} ( ,..., )( ( ) ( , ))
p
p j p k k k
k
E J u u u P u u p C u u u
=
= -
.
(15.3)
is prohibited.
where P(u
1
,...,u
p
) is the joint probability of u
1
,...,u
p
.
It is clear that Eq.(15.3) will be minimized if
0
( ) ( , )
p
k k k
k
C u u u
=
is maximized with regard to the fnal decision u

0
. Therefore, we defne that the
optimum fnal decision u
0
for multi-biometric should maximize
0
( ) ( , )
p
k k k
k
C u u u
=
.
The rationale to do so can be partially illustrated by the following case: if all C
k
(u
k
) are equal to 1,
0
( ) ( , )
p
k k k
k
C u u u
=
will has the maximum value only if u

0
is the same as the majority of all the local
decisions. More specifcally, in a case of personal verifcation where more than half
of the local decisions accept the person as the genuine user,
0
( ) ( , )
p
k k k
k
C u u u
=
will be maximized only if u

0
also takes the accept decision. In a case of personal
identifcation,
0
( ) ( , )
p
k k k
k
C u u u
=
will be maximized only if u

0
takes the same classifcation decision as the majority
of u
1
,
...
,u
p
. Therefore, we can conclude that if all C
k
(u
k
) are equal to 1, the decision
fusion rule on the basis of the maximization of the expected cost as shown in Eq.
(15.3) will be identical to majority voting. As for other C
k
(u
k
), a decision u
k
with a
large weight C
k
(u
k
) has great effect on the fnal decision. This is why the decision
fusion rule presented above is called as the weighted majority rule.
15.2.6.2 How to Set the Weight
Though we show that maximization of
0
( ) ( , )
p
k k k
k
C u u u
=

is prohibited.
can be regarded as the goal of the optimum fnal decision, how to set the weight
C
k
(u
k
) is still not answered. We now will address this issue. We can assign a large
cost to the fusion rule when it fails to follow a local decision of large reliability.
Because the decision made by different biometric traits have different reliabilities,
we use ( , ), 1, 2,...,
k
rel k u k p = to denote the reliability of the decision u
k
made by
using the kth biometric trait. We use REL(k),k = 1,2,...,p to represent the reliability
of the kth biometric trait as a whole.
Because the decision fusion presented in this paper is devoted to authentication,
we may assume that a biometric trait with higher authentication accuracy is more
reliable than the others. A simple way to set REL(k),k = 1,2,...,p is to let it be directly
associated with the authentication accuracy of the kth biometric trait.
As for rel(k),k = 1,2,...,p, we can specify that it should be also large for a reliable
local decision and consequently the fnal decision is affected as much as possible
by this reliable local decision. Hence, We defne that
( , ) ( | ), 1, 2,...,
k j k j k j
rel k u P x u k p = = = = (15.4)
where x
k
stands for the sample of the kth biometric trait. Indeed, Eq.(15.4) takes that
the probability that the local decision u
k
is correct as the reliability of u
k
. Eq.(15.4)
also implies that the higher the probability of the local decision u
k
being correct
is, more strongly u
k
infuences the fnal decision of x
k
.
In order to associate the reliability measures of the biometric trait and the local
decision with the weight C
k
(u
k
) of the weighted majority decision fusion rule, we
can simply defne that
( ) ( , ) ( ), 1, 2,...,
k k k
C u rel k u REL k k p = = . (15.5)
15.2.7 Decision Combination Based on the Confusion Matrix
If classifers are of mutual independency, we can evaluate the error of each classifer
using the confusion matrix and combine classifers based on confusion matrices
(Congalton, 1991). For a multiple class recognition problem with M classes, the
error for the kth classifer e
k
can be represented by a two-dimensional confusion
matrix as shown in Eq.(15.6).
is prohibited.
11 12 1
21 22 2
1 2
...
...
.
.
.
...
k k k
M
k k k
M
k
k k k
M M MM
n n n
n n n
PT
n n n
(
(
(
(
( =
(
(
(
(

, k = 1,...,K (15.6)
where
k
ij
n
denotes the number of the samples from the ith class while classifed into
the jth class by the kth classifer.
The confusion matrix can be clearly illustrated in the case of two-class clas-
sifcation. In this case, the confusion matrix is a 2 by 2 matrix, which contains
information about actual class and predicted class by classifcation decisions of
the classifer. If the four entries of the confusion matrix are a, b, c, d, respectively,
then the meanings of these entries are shown in Table 15.1.
a is the number of correct predictions that the sample is from the negative class;
b is the number of incorrect predictions that the sample is from the positive class; c
is the number of incorrect predictions that the sample is from the negative class, and
d is the number of correct predictions that the sample is from the positive class.
The defnition of accuracy, which is shown in Eq.(15.7), is helpful to show the
classifcation performance of a classifer for the two-class classifcation problem.
Theaccuracy(AC) is defned as the proportion of the total number of predic-
tions that were correct.
a d
AC
a b c d
+
=
+ + +
. (15.7)
Note that the accuracy determined using Eq.(15.7) may not be an adequate per-
formance measure when the number of negative or positive cases is much greater
Table 15.1. Meanings of the entries in the confusion matrix of a two-class classi-
fer
Predicted class by the classifcation decision
Negative class Positive class
Actual
class
Negative class a b
Positive class c d
is prohibited.
than the number of the other cases. For example, suppose there are 500 cases, 490
of which are negative cases and 10 of which is positive cases. If the system classifes
them all as negative, the accuracy would be 98%, even though the classifer missed
all positive cases. However, the following defnitions can be used to provide more
measure for the classifer performance.
Theprecision (P) is defned as the proportion of the predicted positive cases
that were correct.
d
P
b d
=
+
.
The recall or true positive rate (TP) is defned as the proportion of actual
positive cases that were correctly classifed.
d
TP
c d
=
+
.
Thefalsenegativerate (FN) is defned as the proportion of actual positives
cases that were incorrectly classifed.
c
FN
c d
=
+
.
The false positive rate (FP) is defned as the proportion of actual negatives
cases that were incorrectly classifed.
b
FP
a b
=
+
.
Thetruenegativerate(TN) is defned as the proportion of actual negatives
cases that were classifed correctly.
a
TN
a b
=
+
.
We can evaluate belief measure of recognition for each classifer on the basis of
the confusion matrix as follows:
( | ( )) ( | ( ) )
i k i k
Bel x c e x P x c e x j = = , (15. 8)
( ) ( )
1
( | ( ) )
M
k k
i k ij ij
i
P x c e x j n n
=
= =

where
( ) k
ij
n denotes the number of the samples from the ith class while classifed into
the jth class by the kth classifer, i,j = 1,2,...,M. e
k
stands for the kth classifer. e
k
(x)
is prohibited.
= j means that the kth classifer classifes the sample x into the jth class. Indeed, the
belief measure
( | ( ))
i k
Bel x c e x
is used to denote the conditional probability that x
is from the ith class (i.e.
i
x c ) under the condition that the event e
k
(x) = j occurs.
Combining the belief measures of all classifers, we obtain the following fnal
belief measure of the multiple classifer system (Chen, Wang, & Chi, 1997):
1
1 1
( | ( ) )
( )
( | ( ) )
K
i k k
k
K M
i k k
i k
P x c e x j
Bel i
P x c e x j
=
= =
=
=
=

. (15.9)
The fnal classifcation decision can be made as follows: if
1,2,...,
( ) max ( )
i M
Bel k Bel i
=
= ,
then the sample is classifed into the kth class. The belief of the fnal classifcation
decision is
1,2,...,
( ) max ( )
i M
Bel k Bel i
=
= .
Note that one of the limitations of the decision combination based on the confusion
matrix is that it requires mutual independencies among multiple classifers which
do not usually hold in real-world applications (Gee & Abidi, 2000).
15.3 select Ing cl Ass If Iers bAsed on correl At Ions
BEt WEEN CLASSiFiERS
In a number of cases, multiple available decisions are made by different classifers.
In these cases, decision fusion is directly associated with the classifer selection
problem. The reason to select classifers for multi-classifer-based decision fusion
is that the combination of different classifers can result in different fusion perfor-
mance. A proper set of classifers should be robust and be able to generate the best
fusion performance. Among a number of available selection methods such as Q
statistic, generalized diversity and agreement (Kuncheva, Whitaker, Shipp, & Duin,
2003; Partridge & Krzanowski, 1997; Petrakos & Benediktsson, 2001), the degree
of correlation is an index of agreement of classifers. It is commonly admitted that
the dependency among classifers can affect the fusion results. Goebel, Wilson and
Sirohey (1995) selected classifers according to the correlation degree of n different
classifers defned as follows:
f
n
f c f
nN
N N N nN
=
- - +
, (15.10)
is prohibited.
where N
f
stands for the number of samples that are misclassifed by all classifers,
N
c
denotes those samples that are classifed correctly by all classifers and N is the
total number of the classifed samples.
Generally, a smaller correlation degree can lead to better performance of classifer
fusion because the independent classifers can give more effective information.
We can select a set of classifers using the following steps:
Step1:Select an appropriate performance measure as the initial evaluation criterion,
such as the ratio of number of samples classifed correctly to the total samples;
Step2: Find the best classifer as the frst classifer of the set of classifers;
Step3:Calculate the correlation degree between the frst classifer and the other
classifers using Eq. (15.10);
Step4: Select the classifer having the low correlation and add it to the set of
classifers.
Step5: Repeat Steps 3 and 4. Classifer selection is not terminated until a suffcient
number of classifers have been selected. Then the optimal sequence of classifers
can be obtained.
15.4 A cAse study of grou P dec IsIon-b Ased f Ace
r ecogn It Ion
Among the various face feature extraction methods (Chellappa, Wilson, & Sirohey,
1995), the algebraic feature method is a kind of effective method to represent the
face feature. The singular value (SV) feature is a commonly used algebraic feature.
It was used by (Hong & Yang, 1991; Cheng & Yang, 1992) for face image classifca-
tion and obtained promising experimental results. However, they adopted the face
samples with small variations and do not preprocess the image. Hence, when the
samples have more changes in expression, pose and illumination, the SV feature
cant obtain the ideal result. Therefore, in order to improve its classifcation perfor-
mance, we design a new algorithm based on a group decision-making approach. We
frst preprocess the face images by employing the orthogonal wavelet transform,
respectively extract the SV features of four sub-images, and then perform classi-
fcation for each sub-image using the conventional nearest neighbor classifer. We
design a real-time classifer combination algorithm, termed Group Decision-Mak-
ing (GDM), to combine the four classifcation procedures of the four sub-images.
We also compare the combination result of our algorithm with the classifcation
result of the SV feature of the original image, and the combination result of the
conventional mark-counting method.
is prohibited.
15.4.1 singular f eature extraction
Wavelet transforms have been widely applied in the feld of the image processing. As
a powerful tool of signal analysis and processing, a wavelet transform decomposes
a signal into different frequency bands. The bandwidth of the flters is not variant
in the logarithmic scale. This is similar with a model of human vision, which also
has the attribute of logarithmic variation. An image signal may be regarded as a
kind of nonlinear and non-smooth signal.
An orthogonal wavelet transform can not only keep the original information
of an image but also has the attribute of unique decomposition and decorrelation.
We select a normal orthogonal wavelet base, namely Daubechies wavelet base, to
perform discrete wavelet transform for a face image of 64 64 pixels. Figure 15.1
shows an original image and the resultant image of the wavelet decomposition.
Note that the resultant image are indeed four subimages, namely left-up one, right-
up one, left-down one and right-down one, which are respectively approximation,
horizontal, vertical, and diagonal detailed subimages.
An algebraic feature of an image extracted by an algebraic transform or matrix
decomposition is an intrinsic feature of the image. As a kind of algebraic feature,
SV features have the following attributes (Hong & Yang, 1991): (1) they are not
sensitive to the variation of image noise and illumination. (2) They are invariable
to image translation and an image mirror transform.
We frst extract the SV feature from the original image of 64 64 pixels and
obtain 64-dimensional feature vector. We use the conventional nearest neighbor
classifer to classify the obtained feature. Then we respectively apply the above
procedure to the four subimages generated from a wavelet transform. Consequently
we obtain four groups of classifcation results.
15.4.2 group Decision and Combination of Classifcation
r esults
We design a group decision-making approach for effectively combining different
classifers (Jing, Zhang, & Yang, 2003). Generally, a group decision-making ap-
Figure 15.1. Original image and four wavelet subimages
is prohibited.
proach is applied in the decision problem of M schemes and N principles (Bryson,
1996). We will gradually introduce the relational concepts and algorithm.
First, we take the Shannon entropy as the divisibility measurement of classi-
fcation results. This requires us to give the estimate of the posterior probability.
When the number of training samples is small and we use the nearest neighbor
classifer, the posterior probability can be approximately evaluated by the similar
value between the training sample and the testing sample. Generally the smaller
their distance, the larger their similarity. Assume that c is the class number, d
i
is
the distance between the ith class and the testing sample, and u
i
is the estimation
of posterior probability. The object function is defned as follows:
2
1
c
m
i i
i
J u d
=
=
and
1
1
i
c
i
u
=
=
(15.11)
where m is the fuzzy index, m > 1. Solving the above equation to obtain the maxi-
mum value of J, we have
1 1
1 1
2 2
1
1 1
( ) ( )
c
m m
i
i j j
u
d d
- -
=
=

(15.12)
If m = 2 , Eq.(15.12) will be rewritten
2 2
1
1 1
( ) ( )
c
i
i j j
u
d d
=
=

. (15.13)
After computing u
i
using Eq.(15.12), we can calculate the Shannon entropy
value.
The similarity between classifers should also be evaluated. Let x, y be two c-
dimensional classifcation result vectors. We can use s(x,y) = 1 sin(x,y) to stand
for the similarity of two vectors. We present several defnitions as follows.
Defnition 15.1 Suppose that x, y are c -dimensional vectors, then
cos( , )
|| || || ||
x y
x y
x y
,
2
sin( , ) 1 cos ( , ) x y x y = -
,
( , ) 1 sin( , ) s x y x y = -
,
(15.14)
We call a similarity measurement between x and y. It is clear that .
is prohibited.
Defnition 15.2 Suppose that x
d
(d = 1,2,
...
,M) is the classifcation result correspond-
ing to the dth decision-maker and h
d

is the entropy value of the dth decision-maker.
The individual consistency index of the dth decision-maker is defned as follows:
( )
( )
( )
1,
, 1 1
M
d d r d
r r d
I s x x M h
=
| |
| = -
|
\ .
(15.15)
Clearly, measures the average similarity between an individual and the group
divisibility. Defnition 15.2 tells us that the larger is, the more representative the
dth decision-maker is for the group.
Defnition 15.3 The combination result of group decisions is defned as
1 1
( , )
M M
G d d d
d d
x s I x I
= =
=

.
Defnition 15.4 The group consistency index is formulated as follows:
1
( , )
M
G G d
d
I s x x M
=
=
(15.16)
Defnition 15.5 If I
G
, we say that the group is consistent in level ; otherwise
we say it is inconsistent in level . The proper value of can be determined accord-
ing to the specifc details of the application.
The key idea of our approach is as follows: frst, under the constraint of I
G
,
fnd a consistent decision for the whole group. This means either using all classifers
that can provide a commonly supported opinion, or selecting part of them which can
provide an opinion for the majority. Second, perform linear classifers combination
according to I
d
(d = 1,2,
...
,M). The concrete algorithm of group decisiom-making is
described in the following.
Step1. We compute
I
d
(d = 1,2,
...
,M) using Defnition 15.2 and sort them in descending order.
Step 2. Evaluate the group decision-making using
1 1
( )
M M
G d d d
d d
x I x I
= =
=

. (15.17)
is prohibited.
Step2. Calculate I
G
using Eq.(15.16). If I
G
, then x
G
is regarded as the result
of group decision-making; otherwise, go to Step 3.
Step 3. If I
G
, assume that the sort result in Step 1 is
1 2
( )
M
I I , I , , I = .
Defne that
( ) d d d L
t I I = , where I
d (L)
is the average value of all the components of I
ranking behind I
d
, i.e.
( )
( )
( ) 1 2 d L d d M
I I I I M d
+ +
= + + + - . (15.18)
Let
{ } { }
( )
max max
d d d L
d d
T t I I = = . (15.19)
The frst k classifers are selected to form a sub-group that plays the dominant
role in the whole group. Repeat Step 1 using this new group which has k available
classifers. The fnal solution of x
G
is the result of group decision-making.
15.4.3. experiment r esults
We conducted an experiment using the NUST603 face databases (Jin, Yang, Hu,
& Lou, 2001) of Nanjing University of Science and Technology, China. The face
images of this database are from 18 persons each providing 12 captured images of
64 64 pixels. There are signifcant changes in both expression and pose, and small
changes in illumination and the relative distance between the camera and person.
The group consistency level is set to 0.8. In other words, the similarity value
between the vectors is 0.8. Table 15.1 shows the results, including those for the
original images, four group subimages, the Borda count method, and the group
decision-making approach of setting the group consistency level or not.
From Table 15.2, we learn that combining the four group classifcation results can
effectively improve the classifcation performance of the SV feature, using either of
two combination methods. The maximum improvement in accuracy obtained using
group decision-making is 12.6%. Moreover, the group decision-making approach
can obtain better combination results than the conventional Borda count method.
In particular, the group decision-making approach setting the group consistency
level obtains a maximum improvement in accuracy of 7.3% in comparison with
the Borda count method.
is prohibited.
15.5 comments on bI ometr Ic f us Ion At the t hree
l evels
In this and the previous two chapters, we have presented three categories of multi-
biometric fusion strategies i.e. feature level fusion, matching score level fusion
and decision level fusion. A number of studies have shown the effectiveness and
power of the multi-biometric system using these fusion strategies. The advantages
of decision level fusion include easy implementation and a simple way to directly
fuse decisions of different biometric subsystems. However, unlike feature level and
matching score level fusion, decision level fusion does not allow information of
multiple biometric traits to be fully exploited. Indeed, decision fusion for verifcation
fuses only the accept or reject decisions associated with different biometric traits
and decision fusion for identifcation fuses only the class label decisions associated
with different biometric traits. It is commonly considered that fusion at the decision
level does not has the same potential to improve the overall system performance as
Table 15.2. Recognition results of different approaches. I, I1, I2, I3 and I4 respec-
tively stand for right recognition rates of original image, approximation, horizontal,
vertical, and diagonal detailed subimages. B, G1 and G2 respectively denote the
Borda count method, the group decision-making approach, with and without setting
the group consistency level .
Different sample sets
Recognition rate (%)
I
Four subimages
B
G
I1 I2 I3 I4 G1 G2
Training
sample
number
per
class
4
Test class
number
12 78.5 77.6 61.8 67.4 58.3 81.3 84.7 85.4
14 74.7 73.2 58.3 60.7 56.0 79.2 81.6 82.7
16 71.8 71.1 56.3 60.2 54.7 79.7 82.8 84.4
18 70.5 69.4 52.8 58.3 48.2 75.5 80.1 81.2
5
Test class
number
12 80.7 80.1 65.3 68.0 59.7 81.9 84.0 86.8
14 75.6 74.4 61.3 60.1 58.9 78.6 82.1 83.9
16 73.4 72.9 60.4 59.4 57.3 82.3 81.9 83.4
18 71.1 69.9 54.2 56.9 55.0 78.2 79.2 79.6
6
Test class
number
12 82.6 81.8 65.2 70.8 59.9 84.6 88.2 90.3
14 77.2 76.8 60.1 63.7 60.1 79.6 84.5 86.9
16 75.4 74.0 60.7 62.5 59.9 81.8 84.4 86.8
18 75.0 71.3 54.5 58.8 56.3 77.8 80.1 81.9
is prohibited.
fusion at the matching score level or fusion at the feature level. Additionally, simple
decision level fusion rules such AND or OR may have diffculties in simultane-
ously obtaining satisfactory FAR and FRR. Fusion at the feature level fuses mul-
tiple biometric traits at the most early stage and can enable these traits to be fully
exploited for personal authentication. However, fusion at the feature level also has
weaknesses. It may encounter diffculties in properly combining feature vectors
that are generated from different biometric traits and are possibly non-compatible.
They may also have diffculties in the cases where the user does not possess all
the biometric traits, which makes fusion at the feature level not feasible. Fusion at
the matching score level fuses multiple biometric traits at a later stage than fusion
at the feature level and can exploit more information generated from the multiple
biometric traits than fusion at the decision level. It seems that fusion at the match-
ing score level is used most among the three fusion strategies.
Our categorization of multi-biometric technologies into three levels is not meant
to imply that multiple biometric traits should be fused at only one of these three
levels. Indeed, hybrid fusion technologies, such as using multiple strategies or
rules (Aude, Carneiro, & Serdeira, 1999; Scheunert, Lindner, Richter, & Tatschke,
2007; McCullough, Dasarathy & Lindberg, 1996), have received more and more
research interests and achieved encouraging results. Some new methods such as
semantic fusion (Oermann, Scheidat, & Vielhauer, 2006), fuzzy theory (Lau, Ma,
& Meng, 2004) are also being exploited for personal authentication in the feld of
biometric research. Evidential reasoning (Reddy, 2007) has also been introduced
into the multimodal fusion of human computer interaction. Adaptive multi-bio-
metrics is also an attractive example of new methods in the feld of biometrics.
An adaptive multi-biometric system will adaptively determine the weights for
different biometric traits or adaptively select the most suitable biometric traits to
conduct a varying multi-biometric implementation with various environments or
conditions (Hui, Meng, & Mak, 2007). Moreover, adaptive multimodal biometrics
(Veeramachaneni, Osadciw, & Varshney, 2005) can adaptively determine the deci-
sion fusion rule and even threshold values for decision-making. Chatzis, Bors and
Pitas (1999) showed the advantage of using the unimodal reliability information for
fusing multimodal biometrics. The fuzzy clustering algorithm can also be used for
fusing multi-biometric traits at the decision level (Chatzis, Bors and Pitas 1999).
Soft computing for biometrics (Franke, Ruiz-del-Solar, & Kppen, 2002) might
provide the multi-biometric system with new available biometric information. It
is considered that since soft computing has the ability to consider variations and
uncertainty in data, it is suitable for biometric measurements that do not have an
absolute ground truth. Existing approaches to and techniques for multi-biometric
fusion also include multivariate polynomial model, byperbolic functions and neural
networks (Toh, Yau, & Jiang, 2004; Toh & Yau, 2004).
is prohibited.
r eferences
Aude, E., Carneiro, G., & Serdeira, H. (1999). CONTROLAB MUFA: a multi-level
fusion architecture for intelligent navigation of a telerobot. Proc. IEEE International
Conference on Robotics and Automation (pp. 465-472).
Bryson, N. (1996). Group decision-making and the analytic hierarchy process:
exploring the consensus-relevant information bontent. Computers & Operations
Research, 23(1), 27-35.
person authentication. IEEE Trans. System, Man, and Cybernetics, part A, 29(6),
674-680.
Chellappa, R., Wilson, C., & Sirohey, S. (1995). Human and machine recognition
of faces: a survey. Proceedings of the IEEE, 83(5), 705-740.
Chen, K., Wang, L., & Chi, H. (1997). Methods of combining multiple classifers
with different features and their applications to text-independent speaker identi-
fcation. International Journal of Pattern Recognition and Artifcial Intelligence,
11(3), 417-445.
Cheng, Y. Q., & Yang, J. Y. (1992). Aircraft identifcation based on the algebraic
method. SPIE (pp. 298-305).
Congalton, R. G. (1991). A review of assessing the accuracy of classifcations of
remotely sensed data. Remote Sensing of Environment, 37(1), 35-46.
Dop, E. V. (1999). Multi-sensor object recognition: The case of electronics recycling.
Ph.D thesis, University of Twente, Holland.
Fauvel, M., Chanussot, J., & Benediktsson, J. A. (2006). Decision fusion for the
classifcation of urban remote sensing images. IEEE Trans. Geoscience and Remote
Sensing, 44(10), 2828-2838.
Franke, K., Ruiz-del-Solar, J., & Kppen, M. (2002). Soft biometrics: Soft computing
for biometric applications. International Journal of Fuzzy Systems, 4(2), 665-672.
Freedman, D. D. (1994). Overview of decision level fusion techniques for identifcation
and their application. Proc of the American Control Conference (pp.1299-1303).
Gee, L. A., & Abidi, M. A. (2000). Multisensor fusion for decision-based control
cues. Proc. of SPIE on Signal Processing, Sensor Fusion, and Target Recognition
IX (pp. 249-257).
is prohibited.
Goebel, K., Yan, W. Z., & Cheetham, W. (2002). A method to calculate classifer
correlation for decision fusion. Proc. of Decision and Control (pp. 135-140).
algorithms for 3D face recognition. Proc. of International Conference on Pattern
Gokberk, B., Salah, A. A., & Akarun, L. (2005). Rank-based decision fusion for 3D
shape-based face recognition. Proc. of Audio- and Video-based Biometric Person
Authentication (pp. 1019-1028).
Hall, D. L., & Llinas, J. (1997). An Introduction to Multisensor Data Fusion. Pro-
ceedings of the IEEE, 85(1), 6-23.
Hong, L., & Jain, A. K. (1998). Integrating faces and fngerprints for personal iden-
tifcation. IEEE Trans. Pattern Anal. Machine Intell., 20(12), 1295-1307.
Hong, Z. Q., & Yang, J. Y. (1991). Algebraic feature extraction of image for recogni-
tion. Pattern Recognition, 24(2), 211-219.
Hui, H. P., Meng, H. M., & Mak, M. W. (2007). Adaptive weight estimation in
multi-biometric verifcation using fuzzy logic decision fusion. Proceedings of the
IEEE International Conference on Acoustics, Speech and Signal Processing (pp.
501-504).
Jain, A. K., Lin, H., Pankanti, S., & Bolle, R. (1997). An identity-authentication
system using fngerprints.Proceedings of the IEEE, 85(9), 1365-1388.
Jeon, B., & Landgrebe, D. A. (1999). Decision fusion approach for multitemporal
classifcation. IEEE Trans. Geoscience and Remote Sensing, 37(3), 1227-1233.
Jin, Z., Yang, J. Y., Hu, Z., & Lou, Z. (2001). Face recognition based on the uncor-
Jing, X. Y., Zhang, D., & Yang, J. Y. (2003). Face recognition based on a group
decision-making combination approach. Pattern Recognition, 36(7), 1675-1678.
Kittler, J., Hatef, M., Duin, R. P. W., & Matas, J. (1998). On Combining Classifers.
IEEE Trans. Pattern Analysis and Machine Intelligence, 20(3), 226-239.
Kuncheva, L. I., Whitaker, C. J., Shipp, C. A., & Duin, R. P. W. (2003). Limits on
the majority vote accuracy in classifer fusion. Pattern Analysis and Applications,
6(1), 22-31.
Lau, C. W., Ma, B., & Meng, H. M.. (2004). Fuzzy logic decision fusion in a mul-
timodal biometric system. Proceedings of the 8th International Conference on
Spoken Language Processing (ICSLP) (pp. 108-115).
is prohibited.
Li, S. Z., Zhao, C., Ao, M., & Lei, Z. (2005). Learning to fuse 3D+2D based face
recognition at both feature and decision levels. IEEE International Workshop on
Analysis and Modeling of Faces and Gestures (pp. 44-54).
McCullough, C. L., Dasarathy, B. V., & Lindberg, P. C. (1996). Multi-level sensor
fusion for improved, target discrimination. Proceedings of the 35th IEEE Confer-
ence on Decision and Control (pp. 3674-3675).
Niu, G., Han, T., Yang, B. -S., & Tan, A. C. C. (2007). Multi-agent decision fu-
sion for motor fault diagnosis. Mechanical Systems and Signal Processing, 21(3),
1285-1299.
Osadciw, L. A., Varshney, P. K., & Veeramachaneni, K. (2003). Optimum fusion
rules for multimodal biometrics. Multisensor Surveillance Systems: A Fusion Per-
spective, Boston: Kluwer Academic Publishers.
Partridge, D., & Krzanowski, W. (1997). Software diversity: practical statistics for
its measurement arid exploitation. Information and Software Technology, 39(10),
707-717.
Petrakos, M., & Benediktsson, J. A. (2001). The effect of classifer agreement on
the accuracy of the combined classifer in decision level fusion. IEEE Trans. Geo-
science and Remote Sensing, 39(11), 2539-2546.
Prabhakar, S., & Jain, A. K. (2002). Decision-level fusion in fngerprint verifcation.
Oermann, A., Scheidat, T., Vielhauer, C., et al. (2006). Semantic fusion for biometric
user authentication as multimodal signal processing. Proceedings of international
workshop Multimedia content representation, classifcation and security (MRCS)
(pp. 546-553).
Reddy, B. S. (2007). Evidential reasoning for multimodal fusion in human computer
interaction. A thesis presented to the University of Waterloo, Ontario, Canada.
Roli, F., Kittler, J., Fumera, G., & Muntoni, D. (2002). An experimental comparison of
classifer fusion rules for multimodal personal identity verifcation systems. Proc. of
the Third International Workshop on Multiple Classifer Systems (pp. 325-335).
Scheunert, U., Lindner, P., Richter, E., & Tatschke, T. (2007). Early and multi level
fusion for reliable automotive safety systems. Proc.2007 IEEE Intelligent Vehicles
Symposium (pp. 196-201).
Teoh, A. B. J., Samad, S. A., & Hussain, A. (2002). Decision fusion comparison
for a biometric verifcation system using face and speech. Malaysian Journal of
Computer Science, 15(2), 17-27.
is prohibited.
Teoh, A. B. J., Samad, S. A., & Hussain, A. (2004). Nearest neighbourhood clas-
sifers in a bimodal biometric verifcation system fusion decision scheme. Journal
of Research and Practice in Information Technology, 36(1), 47-62.
Tien, J. M. (2003). Toward a decision informatics paradigm: a real-time information
based approach to decision making. IEEE Trans. SMC-C, 33(1), 102-113.
Toh, K. A., & Yau, W. Y. (2004). Combination of hyperbolic functions for multi-
modal Biometrics Data Fusion. IEEE Trans. Systems, Man, and Cybernetics, Part
B, 34(2), 1196-1209.
Toh, K. A., Yau, W. Y., & Jiang, X. (2004). A reduced multivariate polynomial
model for multimodal biometrics and classifers fusion. IEEE Trans. Circuits and
Systems for Video Technology (Special Issue on Image- and Video-Based Biomet-
rics), 14(2), 224-233.
Veeramachaneni, K., Osadciw, L. A., & Varshney, P. K. (2005). An adaptive
multimodal biometric management algorithm. IEEE Trans. Systems, Man, and
Cybernetics, Part C, 35(3), 344-356.
verifcation application. Proc. of 2nd Intl Conf. on Audio- and Video-based Person
Authentication (pp. 188-193).
identity verifcation. Proc.of International Conference on Audio and Video-Based
Zuev, Y. A., & Ivanon, S. (1996). The voting as a way to increase the decision re-
liability. Proc. Foundations of Information/Decision Fusion with Applications to
Engineering Problems (pp. 206-210).
Book Summary 349
is prohibited.
Chapter XVI
Book Summary
Abstr Act
With the title Advanced Pattern Recognition Technologies with Applications to
Biometrics this book mainly focuses on two kinds of advanced biometric recognition
technologies, biometric discrimination techniques and multi-biometrics. Biometric
discrimination techniques are presented in Parts I and II, while multi-biometrics
is described in Part III. While the methods and algorithms described in Parts I
and II are very suitable for biometrics as they take into account characteristics of
biometric applications such as high dimensionality and small sample size, Part III
mainly introduces three kinds of biometric fusion techniques that respectively fuse
biometric information at the feature level, matching score level and decision level
as well as their applications cases. This chapter summarizes the book from a holis-
tic viewpoint. Section 16.1 summarizes the contents of the book and indicates the
relationship between different chapters in each part. Section 16.2 reveals that how
the methods and algorithms described in different parts can be applied to different
data forms of biometric traits. Section 16.3 provides comments on the development
of multi-biometrics.
16.1 content summAr y
In this section we summarize the contents of all the three parts and indicate the
themes of different parts, respectively. While Part I explores several advanced
biometric discrimination technologies such as orthogonal discriminant analysis,
parameterized discriminant analysis, and maximum scatter difference discriminant
is prohibited.
analysis, this part is indeed focused on biometrics with the small sample size (SSS)
characteristic. In Chapter IV, by defning the three SSS strategies (SSS Strategy
one, SSS Strategy two and SSS Strategy three) for solving the SSS problem using
discriminant analysis, we provide clear and detailed description on how a dis-
criminant analysis technique should be properly applied to biometrics with the SSS
characteristic. On the other hand, in Chapter V and Chapter VI we develop novel
discriminant analysis methods applicable to biometrics with the SSS characteristic.
In Chapter V, we address biometric issues with the SSS characteristic by develop-
ing two weighted discriminant analysis methods respectively on the basis of the
nullspace and the range space of the within-class scatter matrix. In Chapter VI, we
address biometric issues with the SSS characteristic by proposing the maximum
scatter difference discriminant analysis method. The most advantage of this method
is that when applied to biometrics the nature of the method allows the SSS problem
to be automatically avoided. This excellent property of the method also allows one
to use discriminant analysis for biometrics with ease.
Part II mainly presents tensor-based biometric discrimination technologies, tensor
independent component analysis, tensor non-negative factorization, tensor canonical
correlation analysis and tensor partial least squares as well as their applications to
biometrics. These techniques allow ideas of conventional methods and algorithms
such as the idea of conventional linear discriminant analysis (LDA) methods to be
applied to new data forms such as two-dimensional matrices. For example, whereas
conventional LDA is just applicable to one-dimensional vector, tensor-based dis-
criminant analysis method enables the discriminant analysis methodology to be
directly applied to two-dimensional matrix data. Indeed, a number of tensor-based
methods and algorithms described in Part II can be viewed as improvements to
conventional methods and algorithms that are workable for only sample data in the
form of vectors as shown in Part I. In addition, when we describe well-known two-
dimensional PCA and LDA respectively from the points of view of tensor PCA and
LDA, we provide readers with a more easy way to understand tensor methods.
Part III primarily presents multi-biometric technologies including feature level
fusion, matching score level fusion and decision level fusion technologies and ap-
plications examples of these technologies. Fundamental concepts and defnitions of
multi-biometrics such as multi-biometric technology taxonomies are also presented
in this part. One of the relationships between this part and the previous two parts is
as follows: The methods presented in Parts I and II can serve as feature extraction
methods of multi-biometrics and then the multi-biometric system can implement
the verifcation or recognition task by integrating the multi-biometric technologies
with the feature extraction result. On the other hand, Part III is also different from
Parts I and II as follows: Part III primarily emphasizes technical aspects of multi-
biometrics, whereas Parts I and II mainly focus on theoretical aspects of foundation
Book Summary 351
is prohibited.
of biometrics, presenting different discriminant analysis methods and algorithms
applicable to biometrics.
16.2 method APPl IcAbIl Ity
Biometric data has three typical representation forms: the vector form, the two-
dimensional image form and the 3D matrix form. This book also has the potential
to provide us with three types of biometric discriminant methods applicable to
biometric traits with different representation forms. The methods presented in Part
I are suited to biometric data in the vector form. This type of methods is referred
to as one-dimensional biometric method. Part II provides a type of methods that is
applicable to the two-dimensional image form and is referred to as two-dimensional
biometric method. These methods include two-dimensional LPP, two-dimensional
PCA and LDA as well as GLRAM, and so on. From Part II, we know that it is also
possible for the tensor-based methods to provide us with a tool to directly analyze and
exploit biometric data in the 3D matrix form. This novel type of analysis techniques
for 3D biometric data can be referred to as three-dimensional biometric method.
The three types of biometric methods mentioned above have different charac-
teristics. For the two-dimensional biometric method, aside the low computational
complexity, it also has the advantage of being able to avoid the SSS problem. This
is because the two-dimensional biometric method frst directly exploits the two-
dimensional biometric data i.e. biometric image to produce the so-called total
covariance matrix, between-class scatter matrix, the between-class scatter matrix,
and so on. Because the dimensionality of these matrices is usually smaller than the
number of biometric samples, they are nonsingular matrices. As a result, the SSS
problem does not exist when one solve the certain equation consisting of these non-
singular matrices. In contrast, a conventional feature extraction method usually has
a high computational complexity and suffers from the SSS problem. This is because
the conventional feature extraction method requires that the biometric sample usu-
ally in the form of biometric image be converted into one-dimensional vector in
advance. Consequently, the corresponding total covariance matrix, between-class
scatter matrix, the between-class scatter matrix have a very high dimensionality that
is larger then the number of biometric samples. A conventional feature extraction
method thus encounters the SSS problem. In this sense, the methods and algorithms
in Part II can well address the high dimensionality characteristic of biometrics.
The methods and technologies presented in different parts of this book have
different merits. Our one-dimensional biometric methods presented in Part I not
only can deal with the SSS problem well but also can obtain a high accuracy. Our
two- dimensional biometric methods presented in Part II can be directly and ef-
is prohibited.
fciently applied to biometric data in the form of two-dimensional matrices and
there is no need to in advance convert the data into the one-dimensional form. As
for the three-dimensional biometric method, it has the following characteristic: It
allows us to directly extract features for 3D biometric samples. This indeed provides
us with an effective way to directly exploiting 3D data for performing personal
authentication.
16. 3 comments on mul t I-bI ometr Ics develo Pment
16.3.1 issues and Challenge
Though a number of biometric and multi-biometric systems have been applied in the
real world, there are still challenge and issues. For example, the following several
issues deserve further investigations. First, it is usually considered that multimodal
biometric traits are independent of each other and multiple biometric traits allows
more information to be exploited by the biometric system, which indeed has been
taken as the base for the multimodal biometric system to produce better performance
than the single biometric system. However, for a certain multi-biometric system,
such as a multi-biometric system of fusing multiple feature extraction algorithms or
multiple classifers using only a single biometric trait, it is likely that the fused data
are not independent. For this type of system, the following issues should be consid-
ered: what is the rationale for a multi-biometric system using dependent biometric
information? How to design a fusion algorithm to properly exploit the dependent
biometric information? Second, if the multi-biometric system is applied in the case
of a large user population identifcation system, there are often more than tens or
hundreds of millions of subject images already enrolled in the matcher databases
and the system has to process more than hundreds of thousands of identifcation
requests. In this kind of case, to establish a multi-biometric system that can meet the
following three requirements will be also a challenge task: a quite high accuracy;
low need of staffng levels to properly operate the system; the system is effcient
enough to serve as an on-line application.
Real-world applications of multi-biometrics should also pay attention to the
following issues:
A. Harsh Environments
In the real world, multi-biometrics may be applied in unfavourable climates or online
or mobile environments. In this kind of environment, the possibility of biometric trait
acquisition failure or the acquisition of low-quality biometric traits makes biomet-
Book Summary 353
is prohibited.
ric-based personal authentication more challenging. Therefore, harsh environments
require that the multi-biometric system have high reliability and stability.
Some special challenges and threats to privacy also arise in the case of network
based biometric applications, as the multiple biometric traits may be simultaneously
exposed through the network. A multi-biometric system may reduce threats to pri-
vacy and provide high feasibility by adopting the following recommendations: both
location information and device characteristics should be protected; ease of use of
the mobile identity management tools and simplifed languages and interfaces for
non-experts should be enhanced; the verifable link between the user and his digital
identity has to be ensured (Please refer to: swami.jrc.es/pages/Conference2006.htm).
Other challenges may be encountered in real-world applications of multi-biometrics
include continuous verifcation necessary for high-security environments in which
the protected resource needs to be continuously monitored for unauthorized use
(Janakiraman, Kumar, Zhang, & Sim, 2005; Kumar, Sim, Janakiraman, & Zhang,
2005; Sim, Zhang, Janakiraman, & Kumar, 2007), performance prediction of the
fusion system given the characteristics of the individual sensors (Wang & Bhanu,
2006), etc.
B. Tradeoff between Security and Convenience
Requiring high-quality biometric trait acquisition can improve system security.
Rigorous acceptance thresholds can also improve security against spoofng be-
cause it reduces the acceptance error. But these means might also reduce the
convenience of the system. In addition, though the use of more multiple biometric
traits can make systems highly accurate, acquisition of the biometric traits is also
more inconvenient for the user and takes a longer time than the acquisition of the
single biometric traits. As a result, the system must make a proper tradeoff between
security and convenience.
C. Choice of Biometric Traits and Fusion Strategies
Usually a choice-based combination of biometric traits produces better performance
than arbitrary combination of biometric traits. In addition, appropriate fusion
strategies are also of signifcance for integrating multiple biometric traits, because
different fusion strategies have their own advantages and disadvantages and are
suited to different situations.
is prohibited.
D. Tradeoff between Cost and System Performance
For multi-biometric systems, high performance requirements usually mean higher
system costs including device costs. Therefore, multi-biometric systems with a high
performance to cost ratio are an important goal.
E. Other Negative Factors
Cultural and race background may also affect degree of user acceptance of the
biometric application. Even if in the case where users are unhappy with interac-
tions that impose a high physical or mental workload, or require them to perform
actions they fnd distasteful (SASSE, 2007, p.79), the application may also be
somewhat infuenced.
16.3.2 Development and t rend of Biometrics
16.3.2.1 Requirements of Large-Scale Applications
Note that early biometric applications usually store and use biometric data locally.
Nowadays, large scale distributed biometric applications, which use biometric data
nationally or internationally, have become more and more common. The biometric
applications operate in more variable and even unpredictable environments and
involve more demographically diverse and larger biometric datasets. It seems that
the future requirements of large-scale identity systems for government use will pay
more emphasis on robustness, scalability, and interoperability (Nadel, 2007).
16.3.2.2 High-Resolution and the 3D Biometrics
High-resolution biometrics and 3D biometrics are two signifcant and recently-de-
veloped biometric techniques. Typical high-resolution biometric systems include
high-resolution fnger identifcation system and signature authentication system,
and so on. The fact that a high-resolution signature authentication system is capa-
ble of documenting a large number of levels such as up to 512 different levels of
pressure and a quite high reading resolution (please refer to web site: http://www.
wacom-asia.com/press/20070427.html) allows the system to capture much more
signature information than a low-resolution system. A high-resolution fngerprint
identifcation system allows fngerprints to be authenticated at three different levels
i.e. pattern level, minutia point level, as well as the pore and ridge contour level
(Jain, Chen, & Demirkus, 2006, 2007). A 1,000 dpi fngerprint scan can simultane-
ously capture information at these three levels. The algorithm is another key point
Book Summary 355
is prohibited.
of this kind of system. The algorithm developed for the high-resolution biometrics
should be able to adequately exploit the much information provided by the system.
The high-resolution system cooperating with a fne algorithm dedicated to the
high-resolution data is able to produce a higher authentication accuracy than the
low-resolution system.
3D biometric techniques acquire 3D-image data of biometric traits (Gkberk &
Akarun, 2006; Gokberk, Dutagaci, Akarun, & Sankur, 2007; Gokberk, Salah, &
Akarun, 2005; Li, Zhao, Ao, & Lei, 2005; Sala, Zappa, & Cigada, 2006; Woodard,
Faltemier, Yan, Flynn, & Bowyer, 2006) and the corresponding systems are therefore
called 3D biometric systems. Examples of 3D biometrics include 3D face (Gkberk
& Akarun; Gokberk, Dutagaci, Akarun, & Sankur; Gokberk, Salah, & Akarun;
Li et al.; Sala, Zappa, & Cigada) and 3D fngerprint. While the high-resolution and
the 3D biometric systems promise higher accuracy, they also suffer from some
problems. First, they usually involve a high device cost. Second, since the system
retains a large quantity of information from the subject, it is necessary for the system
to have a large memory and a high computation performance.
16.3.2.3 Cancelable Biometrics and Biometric Key Generation
Issues relating to identity theft, privacy, and cancelable biometrics also attract
increasing attention of the researchers in the feld of biometrics (Bolle, Connel,
& Ratha, 2002; Uludag, Pankanti, Prabhakar, & Jain, 2004). Usually a real-world
biometric system captures and stores a large-scale of biometric data in a database.
When the identity theft issue occurs in the biometric system, serious issues may be
caused. This is because unlike secret passwords or physical tokens, biometric data
in the traditional biometric system cannot be refreshed or reissued.
The technique of cancelable biometrics can enable biometric data to be re-
freshed and reused. Cancelable biometrics stores a transformed version of the
biometric data and it is able to provide a higher level of privacy by allowing mul-
tiple templates to be generated from the same biometric data (Jin & Connie, 2006).
Different sets of the template can also been stored in different databases. Maltoni,,
Maio, Jain, and Prabhakar (2003) presented the following three principal objectives
of cancellable biometrics:
i. Diversity:No same cancellable template can be used in two different applica-
tions.
ii. Reusability: Straightforward revocation and reissue in the event of compro-
mise.
iii One-way transformation: Non-invertibility of template computation to
prevent recovery of biometric data. Besides that, the formulation should not
deteriorate the recognition performance for sure.
is prohibited.
The goal of diversity and reusability mentioned above requires that different
application use different biometric templates. The biometric system can achieve
this goal in the following way. During enrollment the biometric system stores only
the transformed version of the original biometric sample rather than the biometric
sample itself. Here the transform means a change in the representation of an entity,
where the transformed version may comprise exactly the same information as in
the previous one or may refect a loss or augmentation of information contained
in the original representation. During authentication, the biometric system would
convert the received biometric sample using the same transform and the biometric
matching would be carried out in the transformed space. As different applications
can use different transforms (or different parameters of the same transform), the
template issued for a specifc application can only be used by that application. If
a biometric template is ever compromised, a new one can be issued by using a
different transform. Because such a template does not reveal a users biometric
information, it is called cancelable biometrics (Ratha, Connell, & Bolle, 2001) or
a private template (Davida, Frankel, & Matt, 1998).
Differing from cancelable biometrics, the method of biometric key generation
attempts to consolidate the system security by integrating biometric information
into a cryptosystem (Uludag et al., 2004). It is possible that such a solution is more
secure than cancelable biometrics because of the following factors: The generation
process integrates a private key into the user biometric information in a nice way and
consequently both the cryptographic key and biometric information in the template
are inaccessible to the attacker and the cryptographic key will be released to the
corresponding application upon valid presentation of the user biometric template.
Moreover, since the system does not have to perform the biometric matching at
all, it is no need to access biometric information in the template. The method of
biometric key generation can be exploited for digital content protection, copyright
protection and other similar tasks.
Biometric key generation is also different from traditional cryptosystems. The
traditional cryptosystems such as AES (advanced encryption standard) (National
Institute of Standards and Technology, 2001) accept only identical keys designed for
encryption and decryption. Differing from this, specifc biometric representations
from a subject may vary dramatically. As a result, it is necessary for researchers to
address the issue of producing perfect encryption/decryption performance which
expects that the decrypted message is identical to the encrypted message, given the
imperfect biometric authentication technology (Uludag et al., 2004).
Book Summary 357
is prohibited.
r eferences
Advanced encryption standard (2001). Federal information processing standards
publication 197. National Institute of Standards and Technology (online available:
http://csrc.nist.gov/publications/fps/fps197/fps-197.pdf).
Bolle, R. M., Connel, J. H., & Ratha, N. K. (2002). Biometric perils and patches.
Pattern Recognition, 35 (12), 27272738.
Davida, G. I., Frankel, Y., & Matt, B. J. (1998). On enabling secure applications
through off-line biometric identifcation. Proc. 1998 IEEE Symp. Privacy and
Security (pp. 148157).
algorithms for 3D face recognition. Proc. of 18th International Conference on Pat-
tern Recognition (pp. 1018-1021).
Gokberk, B., Dutagaci, H., Akarun, L., & Sankur, B. (2007). Representation plu-
rality and decision level fusion for 3D face recognition. IEEE Trans. System Man
and Cybernetics, accepted.
Gokberk, B., Salah, A. A., & Akarun, L. (2005). Rank-based decision fusion for
3D shape-based face recognition. Proceedings of the International Conference on
Audio- and Video-Based Biometric Person Authentication (pp. 1019-1028).
Jain, A., Chen, Y., & Demirkus., M. (2006). Pores and ridges: fngerprint matching
using level 3 features. Proc.of 18th International Conference on Pattern Recogni-
tion (ICPR) (pp. 477-480).
Jain, A., Chen, Y., & Demirkus., M. (2007). Pores and ridges: high-resolution
fngerprint matching using level 3 features. IEEE Trans. Pattern Anal. Machine
Intell., 29(1), 15-27.
Janakiraman, R., Kumar, S., Zhang, S., & Sim, T. (2005). Using continuous face
verifcation to improve desktop security. Proc. IEEE Workshop Applications of
Computer Vision (pp. 501-507).
Jin, A. T. B., & Connie, T. (2006). Remarks on BioHashing based cancelable bio-
metrics in verifcation system. Neuro Computing, 69, 24612464.
Kumar, S., Sim, T., Janakiraman, R., & Zhang, S. (2005). Using continous biometrics
verifcation to protect interactive login sessions. Proc. 21st Ann. Conf. Computer
Security Applications (pp. 441-450).
is prohibited.
Li, S. Z., Zhao, C., Ao, M., & Lei, Z. (2005). 3D+2D Face recognition by fusion at
both feature and decision levels. Proceedings of IEEE International Workshop on
Analysis and Modeling of Faces and Gestures, Beijing, Oct. 16.
Maltoni, D., Maio, D., Jain, A. K., Prabhakar, S. (2003). Handbook of fngerprint
recognition. New York: Springer (pp. 301-307).
Nadel, L. (2007). Biometric identifcation: a holistic perspective. SPIE Defense and
Security Symposium Biometric Technology for Human Identifcation IV, April 10,
2007.
Ratha, N., Connell, J., & Bolle, R. (2001). Enhancing security and privacy in bio-
metrics-based authentication systems. IBM Syst. J., 40(3), 614634).
Sala, R., Zappa, E., & Cigada, A. (2006). Personal identifcation through 3D bio-
metric measurements based on stereoscopic image pairs. Proc. of the 2006 IEEE
International Workshop on Measurement Systems for Homeland Security, Contra-
band Detection and Personal Safety (pp. 10-13).
SASSE, M. A. (2007). Red-eye blink, bendy shuffe, and the yuck factor - A user
experience of biometric airport systems. IEEE Security & Privacy, 5(3), 78-81.
ligence, 29(4), 687-700.
Uludag, U., Pankanti, S., Prabhakar, S., & Jain, A. K. (2004). Biometric cryptosys-
tems: issues and challenges. Proc. IEEE (Vol. 92, No. 6, pp. 948960).
Wang, R., & Bhanu, B. (2006). Performance Prediction for Multimodal Biomet-
rics. Proceedings of the International Conference on Pattern Recognition (pp.
586-589).
Woodard, D. L., Faltemier, T. C., Yan, P., Flynn, P. J., & Bowyer, K. W. (2006). A
comparison of 3D biometric modalities. Proc. of Conference on Computer Vision
and Pattern Recognition Workshop (CVPRW06) (2006).
Glossary 359
is prohibited.
Glossary
Preface
PCA,orK-Ltransform: principal component analysis
LDA: Fisher linear discriminant analysis
ICA: independent component analysis
CCA: canonical correlation analysis
PLS: partial least squares
KPCA: kernel principal component analysis
KFD: kernel Fisher discriminant
LMLP: large margin linear projection
OFD: orthogonalized Fisher discriminant
FDS: Fisher discriminant with Schur decomposition
MMSD: multiple maximum scatter difference
DCV: discriminant based on coeffcients of variances
NIST: National Institute of Science and Technology
BSSR1: Biometric Scores Set - Release 1
SSS: small sample size
MNMSE: minimum norm minimum squared-error
MSE: minimum squared-error
ChapterI
AFIS: Automated Fingerprint Identifcation System
ATMs: Automated Teller Machines
DNA: Deoxyribonucleic acid
BSSR1: Biometric Scores Set - Release 1
360 Glossary
is prohibited.
OCR: optical character recognition
NMF: non-negative matrix factorization
CKFD: complete kernel Fisher discriminant
KDDA: kernel direct discriminant analysis
ISOMAP: isometric feature mapping
LLE: locally linear embedding
2DPCA: two-Dimensional PCA
GLRAM: generalized low rank approximation of matrices
LMLP: large margin linear projection
ChapterII
FDC: Fisher discriminant criterion
MSE: minimum squared-error
LSVM: linear Support Vector Machine
FLD: Fisher linear discriminant
FSD: Foley-Sammon discriminant
MSD: maximum scatter difference
N-LDA: nullspace LDA
D-LDA: direct LDA
C-LDA: complete LDA
ChapterIII
FDC: Fisher discriminant criterion
2DPCA: two-dimensional principal component analysis
FSD: Foley-Sammon discriminant
MSD: maximum scatter difference
U-LDA: uncorrelated linear discriminant analysis
OCS: orthogonal complementary space
ChapterIV
OCD: orthogonal component discriminant
EFLD: extended Fisher linear discriminant
EFSD: extended FSD
EOCD: extended OCD
Glossary 361
is prohibited.
ChapterV
PD-LDA: parameterized direct linear discriminant analysis
WN-LDA: weighted nullspace LDA
NN: nearest neighbor
ARR: average recognition rate
WLDA-RWS: weighted linear discriminant analysis in the range of the within-
class scatter
LDA-RWS: LDA in the range of the within-class scatter matrix
ChapterVI
FLD: Fisher linear discriminant
K-L: Karhunen-Love
CV: coeffcient of variation
MCV: minimum coeffcient of variance
ChapterVII
SVD: singular value decomposition
CP: CANDECOMP-PARAFAC
ChapterVIII
No
ChapterIX
FOSDV: Foley-Sammon optimal discriminant vectors
UDV: uncorrelated discriminant vectors
2D: two-dimensional
UIDA: uncorrelated image discriminant analysis
LM: Lius method
DRM: direct recognition method
UK: University of Essex
NHDLDA: non-iterative HDLDA
362 Glossary
is prohibited.
ChapterX
VQ: vector quantization
L2distance: Euclidean distance
PCAI: vertically cantered PCA
PCAII: horizontally cantered PCA
MICA: multilinear independent components analysis
NF: non-negative factorization
ChapterXI
SVM: support vector machines
MHKS: modifed HK algorithm
MatMHKS: matrix-based MHKS
FLS-SVM: fuzzy LS-SVM
STL: supervised tensor learning
MPM: minimax probability machines
TMPM: tensor MPM
DPLS: discriminant partial least squares
HK: Ho-Kashyap
LP: linear programming
LFP: linear fractional programming
QP: quadratic programming
QCQP: quadratically constrained quadratic programming
SOCP: the second-order cone programming
SDP: semidefnite programming
GP: geometric programming
LPM: linear programming machine
DML: distance metric learning
2DPLS: 2D partial least squares
ChapterXII
FRR: false reject rate
FAR: false accept rate
FMR: false match rate
FNMR: false non-match rate
FTAR: failure to acquire rate
Glossary 363
is prohibited.
FTER: failure to enrol rate
ROC: receiver operating characteristic
EER: equal error rate
INCITS: InterNational Committee for Information Technology Standards
ITI: Information Technology Industry Council
FRVT2002: 2002 Face Recognition Vendor Test
ChapterXIII
KDCV: kernel discriminative common vectors
RBF: radial base function
DCV: discriminative common vectors
KDRC: KDCVRBF classifer
KDNC: KDCV and NN classifer
DNC: DCV and NN classifer
POC: phase and orientation code
DC: direct current
ChapterXIV
FFN: feed-forward neural network
ChapterXV
AC: accuracy
P: precision
TP: recall or true positive rate
FN: false negative rate
FP: false positive rate
TN: true negative rate
SV: singular value
GDM: group decision-making
ChapterXVI
AES: advanced encryption standard
364 About the Authors
About the Authors
DavidZhang graduated in computer science from Peking University (1974). He
received his MSc in computer science in 1982 and PhD in 1985 from the Harbin
Institute of Technology (HIT). From 1986 to 1988, he was a postdoctoral fellow at
Tsinghua University and then an associate professor at the Academia Sinica, Bei-
jing. In 1994, he received his second PhD in electrical and computer engineering
from the University of Waterloo, Ontario, Canada. Currently, he is a chair profes-
sor at the Hong Kong Polytechnic University, where he is the founding director
of the Biometrics Technology Centre (UGC/CRC) supported by the Hong Kong
SAR government in 1998. He also serves as visiting chair professor in Tsinghua
University, and adjunct professor at Shanghai Jiao Tong University, Beihang Uni-
versity, HIT and the University of Waterloo. He is the founder and editor-in-chief
of the International Journal of Image and Graphics (IJIG); book editor of the
Springer International Series on Biometrics (SISB); organizer of the International
Conference on Biometrics 2004 and 2006 (ICBA 2004 and ICB 2006); associate
editor of more than 10 international journals, including IEEE Transactions on
SMC-A/SMC-C/Pattern Recognition; chair of IEEE/CIS Technical Committee
on Intelligent System Application; and the author of more than 10 books and 160
journal papers. Professor Zhang is a Croucher senior research fellow, distinguished
speaker of the IEEE Computer Society, and fellow of the International Association
of Pattern Recognition (IAPR).
Fengxi Song graduated in applied mathematics from the Anhui University
(1984). He earned his MSc in applied mathematics from the Changsha Institute of
Technology (1987) and PhD in pattern recognition from the Nanjing University of
Science and Technology (2004). From 2005 to 2007, he was a postdoctoral fellow
About the Authors 365
at ShenZhen Graduate School, Harbin Institute of Technology. Currently, he is
a professor at New Star Research Institute of Applied Technology in Hefei City,
China. He is the author of more than 30 scientifc papers in pattern recognition and
computer vision. He is a reviewer of several international journals such as IEEE
Transactions and Pattern Recognition. His research interests include pattern rec-
ognition, computer vision, image processing, and text categorization.
YongXu was born in Sichuan, China, in 1972. He received his BS and MS degree
at Air Force Institute of Meteorology (China) in 1994 and 1997, respectively. He
then received his PhD in pattern recognition and intelligence system at the Nanjing
University of Science and Technology (NUST) in 2005. From May 2005 to April
2007, he worked at Shenzhen graduate school, Harbin Institute of Technology (HIT)
as a postdoctoral research fellow. Now he is an associate professor at Shenzhen
graduate school, HIT. He also acts as a research assistant researcher at the Hong
Kong Polytechnic University from August 2007 to June 2008. He has published
more than 40 scientifc papers. He is an associate editor of the International Journal
of Image and Graphics, a reviewer of several international journals such as IEEE
Transactions on Systems, Man, and Cybernetics, International Journal of Pattern
Recognition and Artifcial Intelligence and Neurocomputing. His current interests
include pattern recognition, biometrics, machine learning and image processing.
ZhizhengLiang graduated in Department of Automation from TianJin Univer-
sity of Technology and Education in 1999. He received his MSc in Department of
Automation in Shandong University in 2001 and got his PhD in pattern and intel-
ligent system from Shanghai Jiaotong University, P.R.China in 2005. From 2005
to 2007, he is a postdoctoral researcher at Shenzhen Graduate School in Harbin
Institute of Technology. Currently, he is a research fellow at City University of
Hong Kong. His current interests include image processing, pattern recognition
and machine learning.
366 Index
Copyright 2009, IGI Global. Copying or distributing in print or electronic forms without written permission
of IGI Global is prohibited.
i ndex
Symbols
2DCCA 228, 241, 242, 245
2DICA 208, 211, 220, 222, 223
2DLDA 173, 179, 180, 184, 187, 191,
192, 198, 199
2DLPP 180, 182, 184, 185, 193, 195,
196, 198, 199
2DNMF 205, 216, 217, 218, 220, 221,
222, 223
2DPCA 14, 40, 41, 151, 152, 159, 161,
162, 163, 165, 166, 170, 171,
182, 220, 221
2DPLS 228, 247, 248
3D biometric techniques 355
B
between-class scatter matrix 28, 59, 61,
78, 79, 80, 82, 87, 88, 89, 92,
93, 97, 107, 108, 109, 111, 112,
114, 175, 176, 280, 281, 351
biometric data 1, 5, 8, 9, 10, 11, 12,
13, 14, 15, 16, 172, 310, 329,
351, 352, 354, 355
biometric data discrimination 1, 8, 10,
12, 13, 14, 15, 16
biometric discrimination techniques 349
biometric image 10, 294, 351
biometric key generation 356
biometrics 1,22, 150, 202, 226, 228,
254, 255, 256, 257, 260, 262,
263, 265, 266, 267, 268, 270,
271, 272, 274, 276, 291, 294,
301, 303, 310358
biometrics with the SSS characteristic 350
biometric system 1, 2, 5, 6, 7, 8, 9, 11,
15, 226, 254, 255, 256, 257,
258, 259, 260, 261, 262, 264,
265, 266, 273, 275, 305, 306,
307, 311, 317, 318, 323, 326,
328, 329, 330, 331, 332, 343,
344, 346, 350, 352, 353, 355,
356
biometric verifcation 5, 265, 266, 268,
271, 331, 346, 347, 348
C
cancelable biometrics 355, 356, 357
classifer selection 328, 329, 330, 337
coeffcients of variances 14, 28, 106, 131
confusion matrix
330, 334, 335, 336, 337
correlations between classifers 328
Index 367
D
decision level fusion 11, 12, 15, 16,
263, 311, 328, 329, 330, 331,
343, 344, 345, 347, 350, 357
decision level fusion rules 328, 344
decision surface 31, 39
direct LDA 22, 28, 29, 52, 88, 105,
107, 108, 133
discriminant criterion 14, 18, 31, 34, 35,
39, 43, 60, 62, 75, 76, 87, 95,
96, 106, 110, 131, 132, 200
E
Eigenvalue 56
eigenvalue decomposition 79, 82
equal error rate 322
F
face recognition 2, 12, 13, 1538, 41,
51, 52, 53, 54, 55, 56, 57,
69, 76, 84, 104, 105, 106, 107,
108, 119, 120, 126, 131, 132,
133, 150, 170, 171, 172, 200,
202, 203, 204, 223, 224, 225,
228, 266, 268, 291, 301, 302,
303, 304, 328, 330, 346, 347,
357
feature extraction 5, 9, 10, 11, 13, 14,
15, 18, 19, 2030, 37, 40, 52,
54, 59, 61, 62, 63, 65, 66, 70,
72, 73, 7896, 101133, 172,
177, 191, 192, 193, 199, 200,
201, 222, 226, 227, 274, 275,
276, 281, 282, 283, 284, 294,
302, 338, 346, 350, 351, 352
feature level fusion 11, 12, 15, 16, 263,
273, 274, 275, 276, 278, 283,
284, 287, 289, 290, 291, 294,
301, 303, 304, 329, 343, 350
feature weight 89, 90, 91, 93, 95, 96,
98, 99, 100, 103
Fisher discriminant criterion 26, 30, 34,
43, 44, 55, 58, 60, 62, 75, 76,
95, 109, 132
Fisher linear discriminant 9, 12, 13, 26,
58, 59, 61, 107
Foley-Sammon discriminant
26, 41, 58, 59, 63
fusion at the decision level
262, 273, 343, 344
fusion at the feature level
261, 274, 275, 301, 344
fusion at the matching score level
261, 273, 306, 344
fusion process 260, 263, 266, 274, 276
fusion strategies 9, 11, 254, 261, 326,
329, 343, 344, 353
G
Gabor transform 276, 277, 278, 279,
284, 288, 291, 294
generalized Rayleigh quotient
34, 45, 106
group decision 15, 16, 18, 328, 338,
339, 341, 342, 343, 346
H
hybrid fusion technologies 344
L
linear classifer 32, 49, 230
linear discriminant analysis
9, 12, 13, 14, 25, 79, 87, 89,
95, 103, 104, 173, 175, 177,
191, 199, 201, 350
M
matching score level fusion 11, 12, 15,
16, 263, 305, 306, 307, 308,
318, 323, 329, 343, 350
matching score level fusion rules 308
multi-biometrics 1, 9, 12, 15, 254, 256,
260, 262, 263, 265, 266, 276,
323, 329, 344, 349, 350, 352,
353
multi-modal biometrics
9, 260, 263, 266, 276
Multilinear LDA 185, 197
368 Index
N
normalization step 307
Nullspace LDA 87
nullspace linear discriminant analysis
14, 78, 89, 103
O
one-dimensional biometric methods 351
optimal projection direction
26, 31, 34, 43, 44, 45
optimization model 26, 27, 31, 35, 36,
44, 58, 59, 60, 64, 67, 96, 109,
122, 123, 124, 125, 177
orthogonal discriminant analysis 14, 15,
28, 58, 60, 70, 72, 75, 349
P
palm print recognition 6, 12, 15, 291
pattern recognition 1, 5, 13, 25, 27,
28, 32, 37, 55, 56, 60, 61, 78,
107, 119, 132, 228, 250, 251,
282, 313
performance indices 321
performance to cost ratio 354
personal authentication 2, 3, 6, 7, 8, 11,
254, 255, 262, 266, 271, 274,
275, 276, 294, 305, 308, 310,
316, 318, 322, 331, 344, 352,
353
Principal component analysis 150, 319
R
recognition technologies 1, 5, 8, 12, 349
S
Schur decomposition 14, 58, 65, 66, 67
single biometrics 256, 257
small sample size problem 14, 15, 17,
28, 56, 76, 104, 105, 132, 228,
250, 276
SSS problem 13, 60, 61, 350, 351
support vector machine
17, 26, 32, 55, 227, 251
T
tensor 9, 10, 12, 13, 14, 16, 20, 110,
133, 135, 136, 137, 138, 139,
140, 141, 142, 143, 144, 145,
146, 147, 148, 149, 150, 151,
153, 159, 160, 161, 168, 169,
170, 171, 186, 188, 190, 197,
198, 199, 200, 202, 204, 205,
212, 213, 218, 219, 220, 223,
224, 226, 227, 236, 237, 238,
244, 249, 250, 252, 350, 351
U
uncorrelated linear discriminant analysis
37, 41, 126
V
verifcation 2, 5, 8, 9, 17, 25, 257,
258, 259, 262, 265, 266, 267,
268, 269, 270, 271, 273, 275,
298, 300, 301, 303, 306, 310,
324, 325, 326, 327, 329, 331,
333, 343, 346, 347, 348, 350,
353, 357, 358
W
within-class scatter matrix 11, 14, 28,
34, 35, 45, 47, 48, 49, 50, 51,
55, 59, 60, 61, 62, 78, 79, 82,
87, 89, 95, 97, 103, 107, 108,
109, 114, 120, 176, 280, 281,
350

Advanced Pattern Recognition Technologies With Applications To Bio Metrics

Uploaded by

Copyright:

Available Formats

Advanced Pattern Recognition Technologies With Applications To Bio Metrics

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Advanced Pattern Recognition Technologies With Applications To Bio Metrics

Uploaded by

Copyright:

Available Formats

What are some of the main problems in biometric recognition discussed in the book?

What are some of the main problems in biometric recognition discussed in the book?

What are some of the main biometric recognition technologies covered in the book?

What are some of the main biometric recognition technologies covered in the book?

Advanced Pattern

1 d d = + ) whose ith row is the transpose of the vector

90.91% 92.73% 92.73%

is an arbitrary matrix, and S

using FLD, FSD,

in ascending order and

in ascending order and

corresponding to zero eigenvalues.

using the formula

. Let W be the discriminant matrix of N-LDA. It follows

If S is singular, there might exist infnite different directions whose correspond-

6.2.1.3 Discriminant Based on Coeffcient of Variance

. For classifcation, the Forbenious-norms between two mode-n

, which satisfes the constraint in Eq.

under approximate constraints. If

Then the multiplicative rule is:

.Then the multiplicative rule is :

H, they seeked a n d non-negative matrix R and

H RC = , and obtain matrices R and C,

. Although the rank of a matrix

S is a matrix with dimension of (n + 1) (n + 1) and

, one can get

Y are nM 1 vectors. Further, we parti-

Y into M blocks respectively, each block having dimensions of n,

12.4.2 Use and Application of Multi-Biometrics

training samples. Then ,

denote the null space of S

is maximized with regard to the fnal decision u

will has the maximum value only if u

will be maximized only if u

will be maximized only if u

334 Zhang, Song, Xu, & Liang

You might also like