See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/351617220

Face Detection and Recognition System

Conference Paper · May 2021

DOI: 10.1007/978-981-16-3153-5

CITATION READS

1 765

2 authors:

Tanjim Mahmud Juel Sikder

Rangamati Science and Technology University , Bangladesh Rangamati Science and Technology University, Rangamati
37 PUBLICATIONS 218 CITATIONS 27 PUBLICATIONS 132 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Tanjim Mahmud on 15 January 2023.

The user has requested enhancement of the downloaded file.

 Face Detection and Recognition System
Tanjim Mahmud 1 , Sajib Tripura 1 , Umme Salma2 , Jannat Fardoush 3 ,
Sultana Rokeya Naher4 , Juel Sikder1 , and Md Faisal Bin Abdul Aziz5
1
Dept. of Computer Science and Engineering,
Rangamati Science and Technology University, Bangladesh
{tanjim_cse, sikder_juel}@yahoo.com
2
Dept. of CSE, Bangladesh University, Bangladesh
3
Dept. of CSE, University of Chittagong, Bangladesh
4
Dept. of CSE,
University of Information Technology and Sciences, Bangladesh
5
Dept. of CSE, Comilla University, Bangladesh

Abstra ct. Facial feature detection and recognition is widely used in the current
world scenarios and technologies. Almost every smart phone is embedding facial
recognition & detection in security matter like phone unlocking. Another
increasing use of this type of technology is being noticed in Camera’s and social
media’s like snap chat. In this paper, we developed a system/platform in which
facial features containing noise, eyes, mouth, and eyebrows can be detected from
an image. The developed system can also recognized a human subject from two
different input images. The proposed system works in two steps -in the first step,
image segmentation has been employed on more than a few images in order to
enhance the images by using various edge detection techniques and filter out the
noise. In second part, features are extracted from image through HOG
(Histogram of Gradient). The entire facial feature detection and recognition
system was developed with histogram of oriented gradients (HOG), support
vector machine (SVM ) and K nearest neighbor (KNN). The proposed system
gives an accuracy of 96% in detection and recognition of facial features and
human subjects. The projected strategy of procedures offers improved truth once
appeared otherwise in relevance the others mixture of ways in which. The
assessment has been done in Python platform on real images of the Indian face
database.

Keywords: Support Vector M achine (SVM ) . Histogram of Gradient (HOG) . K

nearest neighbor (KNN) . Face detection . Face recognition . Facial features

1 Introduction

The searches to identify human faces accurately from existing image frame and the
specific indiv iduals to which they belong was already well underway and have
reached to a new level in research and development in this modern era. In 2012,
Google’s famous X laboratory announced their achievement in facial recognition,
which has amazed the world [1]. They have managed to identify successfully cats on
YouTube with their developed Artificial Intelligence (AI) model. It stands to be a
great example of how far AI models and mach ine learning has reached where the
system just studied just a few thousand YouTube thumbnails bas ed to preset learning
methods. This achievement is a reflect ion of decades of prior research and
development on facial recognition technology. Just like other forms of bio metric
identifications like vo ice or fingerprint identification, facial recognition is given just
as much priority in various modern day technology, cameras and security system [2].
In fingerprint identificat ion, an available line of the fingerprint is used to compare
among fingerprints. Similarly, various aspects of a face can be used to
recognize/identify a human subject fro m images or video frames. In modern day
facial recognition, software’s are able to detect by a critical threshold of resemblance
among sample image and example patterns, a d irect detection is generally declared.
These sorts of modeling have led to detection of facial features such as eyes, mouth,
and nose to find out quite easy. However, when it co mes to identifying entire face o f a
person from mu ltiple photograph, analysis is quite complex. In spite of the fact that
many imp rovement has been done in area of face detection and recognition for
security, recognizable proof and participation reason, yet there are issues upsetting the
advancement to reach or outperform human level exactness. These issues are variet ies
in hu man facial appearance, for example, fluctuating lighting condition, clamo r in
face images, pose, expression and so forth. In this paper, we propose an innovative
combination of methods employing SVM [3] co mbined with HOG[4] and KNN[5] to
address a portion of the issues obstructing face recognition accuracy in order to
improve the widespread face detection and recognition framewo rk. Accordingly
developed system that have an accuracy higher than 90% {or accuracy of 96%} in
detecting facial features (Nose, Left Eye, Right Eye, mouth and Ear) and
identification/recognition of a human subject from mu lt iple test images .
Consequently, this investigation results show that proposed technique is exact, solid
and robust for face detection and recognition framework that can be essentially
actualized, all things considered. The proposed system was developed in open source
software, which supports various advance languages. Python was the preferred
programming language. The investigation has been completed in Python platform on
real images of the Indian face database.
The construction of the paper is according to the going with. In Sect. 2 g ives an
outline of related works, trailed by the proposed technique, exploratory consequences
of assessment, correlat ion of the outcomes with results announced in related works,
conclusion and references.

2 Related Works
Facial appearance can brought to recognition using well-established methods. These
can be stated by looking behind the past two to three decades of research and
development works. In developing any automated systems to be able to recognize
facial appearance and detect facial features, its appropriateness depends on its
application methods used. However, most co mmonly all approaches in develop ment
of these types of systems basically falls upon appearance of the face or geo metrical
calculation. By using a geometric feature-based approach and maximu m recognition
using template matching as claimed [6], the findings obtained resulted in 90 percent
accurate recognition. Haar-like features are evaluated using another image processing
that produces an immense amount of features[7] and uses the AdaBoost boosting
method [8] to minimize the enhanced heart degenerative tree and rapid obstruction
classifiers, using only simple rectangular Haar like features[7] to have different
results, such as a set of ad-hoc area efficiency. A list of capabilit ies that were terribly
enormous will be used in the introduction of a structure that used those features, and
fro m now on the range of features must be limited to a few essential features obtained
by the boosting technique, AdaBoost [8]. Pentland and Matthew Turk [9] applied
Principal Co mponent Analysis (PCA). Reduction of d imensions of images can also be
done with HOG. Facial features or land marks are crucial in recognition and detection.
The facial landmarks are obtained with HOG remov ing unnecessary noise [10]. Th is
procedure insures that the test/samples an image needs not to be used to its entire
image form.
Navneet Dalal and Bill Triggs [4] stated in their analysis that Histograms of
Directed Grad ient (HOG) descriptors significantly outperform current human
detection feature sets. Small-scale gradients, fine orientation binning, comparatively
coarse spatial binning and high-quality local contrast normalizat ion are all important
for successful results.

Table 1. Taxonomy of some most recent related works

Article Accuracy Rate Techniques

[11] 95% PCA and M DC
[12] 85% ICA and M DC
[13] 86.95% M arkov Random Fields

[14] 92.87% Eigen face

[15] 92.60% Statistical
[16] 95.10% HOG and RVM

Table 1. Illustrates the taxonomy of most recent face recognition methods , in light of
their technique and accuracy rate using AT&T datasets [17]. Gabor [18] channels can
misuse remarkable visual properties, for examp le, spatial localizat ion, orientation
selectivity, and spatial frequency qualities. Gabor [18] standards are heartless against
modifications such as lighting, posture and appearances, given these enormous
shortcomings and its excellent achievement in face recognition, and the fact that
Gabor [18] alteration is not strictly meant for face recognition. The formu la for
change is predefined, rather than learned, in the practice o f teaching. Additionally
PCA [11] and LDA classifier thin k about global features while LBP and Gabor
classifier think about local features [18].

3 Proposed M ethodology
Fig. 1 The technique steps are shown to summarize the following primary portion: :
Firstly, has been applied Histogram of oriented gradients (HOG), consequently apply
the classifier: Support Vector Machine (SVM), finally apply the K Nearest Neighbor
(KNN).
3.1 Histogr am of Or iented Gr adients (HOG)
Basically, for features derived fro m an image, HOG is used. Feature ext raction was
started at the starting point and derived functions were created. At a point where the
data of an algorith m is too huge to actually be handled and it is accused of being
replicated, it must be mod ified into a reduced function structure. In the field of visual
recognition and machine vision, HOG is an attribute identifier.

Fig. 1. Block diagram of the developed system

The basic principle behind the histogram of the directed gradient descriptor is that
the dispersion of strength gradients or edge directions will reflect the appearance and
form of the nearby item within a picture [4]. The image is isolated into tiny linked
regions called cells and a histogram of d irected gradients is organized fo r the pixels
within each cell. For these histograms, the reference is the descriptor. The nearby
histograms can be normalized for better results by measuring a proportion of the
amp litude over a wider area of the image, called a block [4]. In order to standardize
all the cells, th is power is then included within the square. This standardization
combines variat ions in enlightenment and shadowing into a larger range. Over other
descriptors, the HOG descriptor has a few favourable conditions, such that it operates
on surrounding cells and is invariant to geometric and image metric changes with the
exception of the object's location. It would simp ly turn up in broader reg ional regions.
Analysing fine align ment and good photographic metric standardization as coarse
spatial areas frequently causes people on foot to ignore the development of the human
body as they expect a role that is usually upstanding. This way, in pictures, the HOG
descriptor is better suited to human positions [10].

3.2 Classifier : Suppor t Vector Machine (SVM)

Technologies for facial recognition are, unquestionably, still an area of study at work.
The algorith m of a support vector mach ine developed by Vapnik(1995) . It has been
associated with different accounts of difficult ies with recognition. In this article, as a
classifier, we use the assistance vector machine (SVM). A Support Vector Machine
(SVM ) is a classifier formally embraced by an insulating hyperplane t hat is
discriminating. When its allocated specified training results are spread, the equation
generates an optimal hyperplane that organizes new models [3], called supervised
learning.
The in itial SVM is a linear binary classifier, which is useful for t wo -class
classification problems[16]. There are very strong relat ions between the SVM and the
neural network. Using the sigmoid kernel function, the SVM model is analogous to a
two-layer perceptron neural network. The goal of SVM modeling is therefore to
identify the optimal hyperplane that isolates vector bunches in such ways that on one
side of the plane are cases with one objective variable classificat ion and on the other
side of the plane are cases with the other classification [3].

3.3 K Near est Neighbor (KNN)

Finally, a strong and simp le classificat ion algorith m as KNN [5] was introduced to
smoothen the process of detection and recognition of facial features. It happens to be
a vital classificat ion algorithm of machine learning in this work. KNN is also a
popular classification algorith m which is widely used in fingerprint detection and
recognition and currently is one of the most suitable algorithms to be used for facial
feature detection and recognition. Normally with this classifier algorith m train ing data
goes through the algorithm and classified coordinates into clusters identified by an
attribute.

4 Experimental Results

At first over 344 subjects human images were co llected. Each subject contains 5
images. Subjects were of various age groups fro m infant child to adult over the age of
50. We proceeded into the experiment by first select test sample images and sized
them to a dimension of 969 pixels.

Fig.2. (a) Hyperplanes to divide into classes; (b-c) Test images with sized dimensions in pixel

As the test, images were prepared and then it goes through proposed developed
system as shown in Fig. 2(b-c). At first only detection was concerned, checked and
accuracy of it was calculated based on the output of detection the proposed system
gave output. At first the test input images goes through HOG for feature descriptor,
noise cancellation. It eliminates unnecessary portion of the image to increase higher
accuracy in detection. It normally converts the images of size width and height which
initially is in vector and finally conversion into length n. As the HOG stages in
completed the descriptive images goes through a comb ination of SVM and KNN.
Here the dataset often decades the fate of the algorith m. At first the KNN go es
through the entire test dataset image and separates them into selective classification
set as parameter of detection according to indiv idual facial features. The parameters a
mainly the eyes, nose, mouth, and eye brow. Further analysis of SVM provides
detections results of facial features. Facial features are identified using red dotted line
all around the particular facial feature selected from designed user interface as shown
in Fig. 3.

Fig. 3. Detection- (a) M outh; (b) Nose; (c) Right Eye

The detection accuracy by developed system was always around 96% every time
we had experimented the developed system. Now co ming to recognition of a hu man
subject from two images Image preparation was done just as done for detection. SVM
and KNN analy zed the input images after it goes through HOG. A threshold value of
67000 in dpi was set which is the hyperplane in SVM port ion as shown in Fig. 2(a).
Anything above that value found fro m the analysis of the co mbination of SVM and
KNN results in that the both input images are not of the same human subject. If the
results scores below the preset threshold value then developed system results that the
recognition among both image is a success and the inputted imag e is of the same
human subject. Here we design user interface results are in plain text. For success in
recognition “Both of the same person”. For detection in not of the same person “They
are not same” as shown in Fig. 4.

Fig. 4. Result-(a). Face recognition; (b). Face recognition for not same person; (c) Face
recognition for same person of the person being of different
For facial feature detection at first face model based on each individual features such
as eyes were the main concern. The coordinates for the center of individual features
were landmarked. Essential features are explained in equation (5)
Here De represents distance between both eyes, which is needed in identifying left eye
to right eye.

Fig. 5. Geometric face model focusing on eye

Gx and Gy , respectively, are the horizontal and vertical co mponents of the dpi
transformation as shown in Fig. 5. For standard scenario, a window size is prefixed
here with a dimension in range between 128 x 144 and 456 x 550 depending if the
image dimension is not always pre-sized to the standard size. Here descriptor values
for each pixels over 8x8 blocks quantized into 9 b lock sizes. Each block represents a
directional angle o f g radient and value in that block. Further Histogram generates
with normalized over a 16x16-b lock size. SVM and KNN are introduced here to go
through analysis here.

∑| |

Here Xi is p ixel of the first input image with which the recognition is done with
whereas for Yi holds pixels of the other image. Fro m these equation results in total
difference which generates a value of dpi. The threshold value then is compared with
total difference. This recognition is done in this system.

Table 2. Results of Face detection

Total faces Detection False detection Average processing M ethods

rate rate time
276 88.3% 2.5% 5.1s AdaBoost+LBP

305 89.5% 2.89% 5.2s

276 92.3% 2.1% 4.6s AdaBoost+Haar

305 94.6% 2.15% 4.8s

276 96.89% 1.3% 4.2 s Proposed M ethod

305 98.1% 1.5% 4.5s

 Table 3. Face Recognition Results

Total Faces Recognition Rate False Acceptance Rate M ethods

344 72.6% 4.1% PCA
344 75.7% 3.9% LDA
344 81.3% 2.6% PCA+LDA
344 83.2% 2.4% LBP
344 92.1% 1.7% Gabor
344 96% 1.5% Proposed M ethod
Tests are conducted on actual Indian face database images [19] and have a higher
outcome for the test collection in terms of 96% accuracy score identification. The
suggested methodology feature extraction technique compares and considers the
accuracy with wh ich the methodology feature extract ion discussed provides greater
precision as seen in Table 2 and Table 3 relat ive to the other feature extraction
technique with the machine-learning algorithm.

5 Conclusions
Face detection and recognition system has planned to solve the issues of existing
manual frameworks. The framework performs acceptable in various pos e and
expression. The system has also been developed to test the face recognition and
recognition techniques that are used as a benchmark fo r current work. Local dat abases
have been used for this purpose to compare a few unifo rmly executed procedures over
various samples, thereby performing different methods arbitrarily depending on the
normal execution of test outcomes. The results of facial recognition and identific ation
strategies on the season, as seen in Table 2 and Table 3, respectively. In current
framework revealed moderately well yet it has a lot of false identification than
developed system. Facial recognition based on HOG feature extraction wo rking very
well in developed system. Developed system g ives 96% accuracy rate on facial
recognition. According as compared to the others techniques our methodology
provides better accuracy. Developed system is also memory efficient because SVM
uses a subset of training points in the decision function. On the other hand, if the
dataset has more noise like target classes are overlapping then the SVM classifier will
not work accurately. In future real time facial recognition can be done by having help
from this system.

References

1. https://www.wired.com/2012/06/google-x-neural-network
2. Rozeha A. Rashid, Nur Hija M ahalin, M ohd Adib Sarijari, Ahmad Aizuddin Abdul Aziz,
Security System Using Biometric Technology: Design and Implementation of Voice
 Recognition System (VRS), Proceedings of the International Conference on Computer and
Communication Engineering, Kuala Lumpur, M alaysia M ay 13-15,(2008)
3. Joachims, T. (1999, June). Transductive inference for text classification using support vector
machines. In Icml (Vol. 99, pp. 200-209).
4. Dalal, N., & Triggs, B. (2005, June). Histograms of oriented gradients for human detection.
In 2005 IEEE computer society conference on computer vision and pattern recognition
(CVPR'05) (Vol. 1, pp. 886-893). Ieee.
5. Imandoust, S. B., & Bolandraftar, M . (2013). Application of k-nearest neighbor (knn)
approach for predicting economic events: Theoretical background. International Journal of
Engineering Research and Applications, 3(5), 605-610.
6. Senior, A. W., & Bolle, R. M . (2002). Face recognition and its application. In Biometric
solutions (pp. 83-97). Springer, Boston, M A.
7. M ita, T., Kaneko, T., & Hori, O. (2005, October). Joint haar-like features for face detection.
In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1 (Vol. 2,
pp. 1619-1626). IEEE.
8. Talele, K. T., Kadam, S., & Tikare, A. (2012). Efficient face detection using Adaboost. In
IJCA Proc on International Conference in Computational Intelligence (Vol. 10).
9. Turk, M . A., & Pentland, A. P. (1991, January). Face recognition using eigenfaces. In
Proceedings. 1991 IEEE computer society conference on computer vision and pattern
recognition (pp. 586-587). IEEE Computer Society.
10.Ghinea, G., & Kannan, R. (2014). Suresh kannaiyan,”. Gradient -Orientation-Based PCA
Subspace for Novel Face Recognition” in IEEE Access, 2, 2169-3536.
11.M ondal, S., & Bag, S. (2017). Face recognition using pca and minimum distance classifier.
In Proceedings of the 5th International Conference on Frontiers in Intelligent Computing:
Theory and Applications (pp. 397-405). Springer, Singapore.
12.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 transactions on
pattern analysis and machine intelligence, 26(1), 131-137.
13.Huang, R., Pavlovic, V., & M etaxas, D. N. (2004, August). A hybrid face recognition
method using markov random fields. In Proceedings of the 17th International Conference on
Pattern Recognition, 2004. ICPR 2004. (Vol. 3, pp. 157-160). IEEE.
14.M a, Y., & Li, S. (2006). The modified eigenface method using two thresholds. International
Journal of Signal Processing, 2(4), 236-239.
15.Rabbani, M . A., & Chellappan, C. (2007). A different approach to appearance–based
statistical method for face recognition using median. International Journal of Computer
Science and Network Security, 7(4), 262-267.
16.Karthik, H. S., & M anikandan, J. (2017, October). Evaluation of relevance vector machine
classifier for a real-time face recognition system. In 2017 IEEE International Conference on
Consumer Electronics-Asia (ICCE-Asia) (pp. 26-30). IEEE.
17.AT&T Database of Faces. :AT&T Laboratories Cambridge.
https://www.cl.cam.ac.uk/research/dtg/attarchive/facedatabase.html(2002)
18.Wiskott, L., Krüger, N., Kuiger, N., & Von Der M alsburg, C. (1997). Face recognition by
elastic bunch graph matching. IEEE Transactions on pattern analysis and machine
intelligence, 19(7), 775-779.
19.https://ieee-dataport.org/keywords/indian-face-database

View publication stats