Robust face recognition using wavelets and neural

networks
Yesenia Sánchez Loredo Rubén Machucho Cadena
Polytechnic University of Victoria Polytechnic University of Victoria
Victoria, Tamaulipas, México. Victoria, Tamaulipas, México.
[email protected] [email protected]

different color model is used and variations in lighting are

Abstract— For humans recognize a person looking at his face present this will result in a poor performance. On the other
is a trivial and daily task that brain performs in seconds and hand, appearance based methods represent optimal feature
almost ever with high precision. Nevertheless, even when this points that can represent the overall facial structure, but with
task seems easy, face recognition remains as a challenge when the disadvantage of being computationally expensive.
implemented in a computer. In this article, a new methodology
The next stage in the flow of a recognition system is the
for accomplish two of the three stages involved in the design of an
automatic face recognition system is presented: for feature
classification, which aims to assign a label a test pattern [8].
extraction stage the discrete wavelet transform is used, and for In the case of face recognition system, the aim is to get the
the recognition stage a personalized multilayer neural network is ID of the person registered on the system database and against
proposed. Experiments were performed using 300 images from which the test image has a higher degree of similarity.
the public database of faces Faces94. The results show a With the goal of carrying out the two steps mentioned
recognition rate of 95%. This demonstrates that the proposed above, this work proposes the use of discrete wavelet
approach is a feasible and effective solution to the face transform as a robust feature extraction tool, due to its
recognition problem. invariance to lighting issues, facial expression, scale and pose,
and the design of a custom multilayer network for
Keywords— Face recognition; neural networks; wavelets classification step.
The paper is organized as follows: section 2 presents
I. INTRODUCTION essential concepts of discrete wavelet transform and its
Face recognition is a research topic that in last years has application in 2D signals, section 3 addresses topics related to
been in the spotlight, due to this technology allow to identify artificial neural networks. Section 4 presents the proposed
a person with more accuracy than traditional systems, which methodology and Section 5 shows the results using a set of
are based in passwords, PIN's magnetic cards, or a test images of the Faces94 public database. Finally, in Section
combination of these. The disadvantage is that this elements 6 we give the conclusions about the work done.
can be transferred, damaged, lost, forgotten or stolen [1,2]. At
II. DISCRETE WAVELET TRANSFORM
this point, biometric systems appear like a robust option for
recognize people facing the problems mentioned. The implementation of the discrete wavelet transform
Among biometric treats that could be used for recognize a (DWT) is based on the use of filter banks for analysis and
person, the face is a interesting characteristic to be studied due reconstruction of signals [9].
its properties of low intrusivity, naturalness, easy use and The DWT uses an orthonormal wavelet family; ie, the
operability even without the voluntary cooperation of the wavelets are orthogonal and normalized to have unit energy
person that we are trying to identify [3, 4]. [10]:
Generally, a face recognition system consists of the
j/2
following stages [5]: image acquisition, detection and
j ,k (t ) 2 (2 j t k ) (1)
segmentation of face, and finally, recognition or classification
step.
In this stages, the feature extraction module is responsible Where j and k are integers that scale and dilate the
for transforming the input data into a set of representative mother wavelet to generate discrete wavelet family.
values and useful as possible for further processing in The practical application of the DWT is performed
subsequent stages of the system [6, 7]. implementing filter banks, whose main idea is to separate the
Among the techniques that have been presented to extract signal into different frequency bands [11]. An example of a
features, template-based methods are easy to implement but signal X(n), which is decomposed into approximations aj(n)
do not represent the overall facial structure; Color and details dj(n) when we apply over this the high-pass filters
segmentation based methods use a color model to detect faces Hj(n) and low-pass filters Gj(n) is shown in Fig. 1, where n is
and morphological operations for detect features, so that if a an integer number and j ={1,2,3…k} is the desired
decomposition level. The union of the two filter outputs Histogram equalization. This activity allowed to
contains the same frequency content that the original input maximize the contrast of the images used in
signal, however, the amount of data is duplicated. Therefore, experiments, with the aim of improving their
decimation by a factor of two, represented as ↓2 is applied to quality and get more accurate results at the stage
the analysis filter bank output. The filter bank can be of face detection and segmentation. In Fig. 2
expanded to an arbitrary level depending on the desired some examples of images to which we applied
resolution. histogram equalization are shown.

Fig. 1. Wavelet decomposition tree to 3 levels.

A. Two dimensional DWT

The two dimensional wavelet transform can be applied to
images, resulting in a matrix of coefficients known as wavelet
coefficients. Its implementation is performed by applying the Fig. 2. Histogram equalization over the images used in experiments.
transform to the rows and columns of the image, which are
seen as one-dimensional signals; this generates four types of Detection and segmentation of the face. We use
coefficients or sub matrices: approximation matrix (LL), the detection algorithm proposed by Viola-Jones,
horizontal details matrix (LH), vertical details matrix (HL) due to their ability to detect effectively and in
and diagonal details matrix (HH) [12]. real time objects of interest within an image with
high accuracy [16]. In Fig. 3 some examples of
III. ARTIFICIAL NEURAL NETWORKS images on which we applied the algorithm are
Neural networks are a robust and widely technique used in shown.
image segmentation and classification problems. Its structure
is essentially a collection of interconnected neurons that are
executed in parallel to perform the task of learning [13, 14].
Its application is given in areas such as: interpreting visual
scenes, speech recognition, control strategies, etc..

Among the major architectures, the multilayer perceptron

is one of the most widely used, because it allow to solves real
problems, in addition to its easy use and application.
Regarding the training, it can be carried out using various
algorithms, where the backpropagation technique is the most
widely used, but this has the disadvantage of being very slow
for its application to real problems. Because of this, there have
been some optimizations to the original backpropagation
algorithm, among which the Levenberg-Marquardt method
stresses, since this achieves a faster convergence to a result, in
contrast to any other existing optimization [15].
IV. PROPOSED METHOD
Fig. 4 shows the block diagram of the method proposed. Fig. 3. Detection and segmentation of the face using Viola-Jones algorithm.
As can be seen, before the step of feature extraction is
desirable to apply some pre-processing techniques, with the Normalizing face size. This activity was carried
goal to improve quality of the images and achieve higher out using the bilinear interpolation method, in
recognition rates. The activities carried out at this stage were: order to adjust the face previously detected and
segmented to a predefined size.
 Fig. 5. Proposed sampling method for sort and divide the approximations
coefficients obtained after application of the transform wavelet.

The proposed sampling technique involves taking only the

coefficients shaded with deeper color than the color of the
background region. This should be carried out for each of the
four regions. Finally, in the current example we can see that of
the total of 100 data constituting each region, after applying
the sampling, only 25 will remain.
That is, of the total of 400 characteristics obtained with
the wavelet transform, only 100 will be considered for
submission to the classifier. Fig. 6 shows the final feature
vector.
Fig. 4. Block diagram of the proposed methodology.

A. Feature extraction
Below is the algorithm proposed in the feature extraction
stage:

1. Take the picture I resulting from the previous

activity and set the number of decomposition levels Fig. 6. Feature vector to represent each face after applying the sampling.
to J.

2. Let F be the wavelet filter used for the C. Classifier design

decomposition. For classification step a personalized multilayer neural
network is proposed. Its architecture is shown in Fig. 7.
3. Apply the discrete wavelet transform to I, using the
low-pass and high-pass filters obtained from F, as
many times as directed by J.

4. Take the approximation coefficients and divide it into

4 sets of equal size, discarding the detail coefficients.

5. If the number of features in the sets exceeds the

amount of 25, proceed to perform sampling on each
set.
B. Sampling technique
Below, is an example that shows how to apply the proposed
sampling method: Setting the size of the face detected at 80 x
Fig. 7. Proposed personalized neural network architecture for classification
80 pixels and applying the wavelet transform using two levels stage
of decomposition, the size of approximation coefficients
region will be 20 x 20 pixels. The coefficients in this region The network has 5 layers. The layers 1, 2, 3 and 4 are
will be sorted and grouped into four regions as shown in Fig. 5 inputs layers, and their responsibility is to receive the input
feature vector. The layer 5 gives the id of the picture in the
database with which a test image is more similar.
The configuration network parameters are listed below:
 Layers 1, 2, 3 and 4 use logsig activation function Number of Recognition Rate
Decomposition using
and three neurons in each of these layers. neurons in
Level Wavelet Training Test
layers 1, 2
Layer 5 receives the output of layers 1, 2, 3 and 4, patterns patterns
3 and 4
and provides the final output of the network. This 4 100% 90%
layer uses a purelin activation function.
The training method is Levenberg-Marquardt.
The network error is set to 0.01. VI. CONCLUSIONS
We presented a methodology for achieve automatic face
The main contribution of this network architecture is that, recognition, based on the use of discrete wavelet transform
compared to a fully-connected multilayer network, the and artificial neural networks.
proposed design provides less weights to adjust, which directly
achieves a reduction of the computational cost required for the The conducted research proved that the combination of
training of the network since the amount of free the wavelet Bior 1.3 and the third decomposition level gives a
parameters (weights) was decreased. recognition rate of 95%, which demonstrated that use of the
wavelet transform provides excellent tool for image
decomposition and textures description.

V. IMPLEMENTATION AND RESULTS Results also allow to validate that the proposed neural network
The proposed methodology was implemented in a desktop architecture provides recognition rates comparable to those
computer with an Intel Core i7 2.4 GHz, 12 GB of RAM and obtained with the use of fully-connected multilayer networks;
operating system Windows 8, 64-bit. The development typically used in face recognition research, but with the
platform used was MATLAB. advantage that the number of weights to adjust in the proposed
The tests were performed by taking 300 images of network is lower, which reduces the computing burden
Faces94 database corresponding to 15 different people, each necessary for training the net.
person has 20 shots; where 16 samples were used for training
and the remaining 4 to measure the generalization capability Finally, it is worth mentioning that results obtained with
of the network. the proposed methodology enable testing with other faces
Experiments using different decomposition levels, databases, and the use of various classifiers with the purpose
wavelet basis and neurons in layers 1, 2, 3 and 4 were of comparing the recognition rates offered by each of these.
performed. The results obtained are shown in Table I, where
we can observe that the best results were obtained using the REFERENCES
third level of decomposition, the wavelet Bior 1.3 and 4
neurons in layers 1, 2 , 3 and 4 of the neural network. [1] S. A. Jassim, Wavelet–Based Face Recognition Schemes, Face
Recognition, Milos Oravec (Ed.), ISBN: 978-953-307-060-5, InTech,
TABLE I. RESULTS USING DIFFERENT LEVELS OF DECOMPOSITION AND 2010.
BASES WAVELETS [2] M. Ruiz-Echartea, Sistema de identificación automática de huella
dactilar latente, Tesis de Maestría, Universidad Politécnica de Victoria,
Number of Recognition Rate 2011.
Decomposition using
neurons in [3] A. Jain and S. Li, Handbook of Face Recognition, Springer-Verlag, New
Level Wavelet Training Test
layers 1, 2 York, Inc., Secaucus, NJ, USA, 2005.
patterns patterns
3 and 4 [4] D. González, Sistemas de reconocimiento facial, DIntel, vol. 2, no. 1,
Daub 4 3 100% 87% pp. 118-119, 2009.
[5] A. Jain, A. Ross, and K. Nandakumar, Introduction to biometrics,
4 100% 92%
Springer-Link : Bucher, Springer US, 2011.
Bior 1.3 3 100% 77% [6] J. Leiva-Murillo, Extracción de características mediante criterios
2 basados en teoría de la información, Universidad Carlos III de Madrid,
4 100% 80% 2007.
Coif 5 3 100% 72% [7] L. Introna and H.Nissenbaum, Facial recognition technology: A survey
of policy and implementation issues, Reporte Técnico, Report of the
4 100% 77% Center for Catastrophe Preparedness and Response, New York
University, 2009.
3 100% 86% [8] A. Jain, R. P. W. Duin, and J. Mao, Statistical pattern recognition: a
Daub 4 review, Pattern Analysis and Machine Intelligence, IEEE Transactions,
4 100% 90% vol. 22, no. 1, pp. 4-37, 2000.
3 Bior 1.3 3 100% 85% [9] R. Merry, Wavelet Theory and Applications. A literature study, 2005.
[10] R. A. González-González, Algoritmo basado en wavelets aplicado a la
4 100% 95% detección de incendios forestales, Universidad de las Américas Puebla,
Coif 5 2010.
3 100% 89%
[11] B. García-Meléndez, E. Mancilla-Ambrona, and R. Montes-Fraile, [13] J. Collantes-Duarte, Predicción con redes neuronales: Comparación con
Optimización de la transformada wavelet discreta (DWT). Universidad las metodologías de Box y Jenkins, Universidad de los Andes, 2001.
Complutense de Madrid, 2005. [14] J. M. Valls, Redes de neuronas. Perceptrón multicapa. Universidad
[12] Jagadeesh, H. S.; Babu, K. Suresh and Raja, K. B., Face recognition Carlos III de Madrid, 2007.
based on DWT and DBC, Computer Communication and Informatics [15] M. Hagan, H. Demuth, and M. Beale, Neural Network Design. Electrical
(ICCCI), 2012 International Conference on , vol., no., pp.1,6, 10-12 Jan. Engineering Series, Pws Pub., 1996.
2012.
[16] P. Viola and M. J. Jones, Robust real-time face detection, Int. J.
Comput.Vision, vol. 57, pp. 137-154, May 2004.