International Journal of Signal Processing, Image Processing and Pattern Recognition

Vol. 2, No. 2, June 2009

51

Face Recognition Using Particle Swarm Optimization-Based
Selected Features

Rabab M. Ramadan and Rehab F. Abdel - Kader
*

Electrical Engineering Department, Faculty of Engineering - Port-Said, Suez Canal
University, Port Fouad 42523, Port-Said, Egypt
[email protected], [email protected]

Abstract

Feature selection (FS) is a global optimization problem in machine learning, which
reduces the number of features, removes irrelevant, noisy and redundant data, and results in
acceptable recognition accuracy. It is the most important step that affects the performance of
a pattern recognition system. This paper presents a novel feature selection algorithm based
on particle swarm optimization (PSO). PSO is a computational paradigm based on the idea
of collaborative behavior inspired by the social behavior of bird flocking or fish schooling.
The algorithm is applied to coefficients extracted by two feature extraction techniques: the
discrete cosine transforms (DCT) and the discrete wavelet transform (DWT). The proposed
PSO-based feature selection algorithm is utilized to search the feature space for the optimal
feature subset where features are carefully selected according to a well defined
discrimination criterion. Evolution is driven by a fitness function defined in terms of
maximizing the class separation (scatter index). The classifier performance and the length of
selected feature vector are considered for performance evaluation using the ORL face
database. Experimental results show that the PSO-based feature selection algorithm was
found to generate excellent recognition results with the minimal set of selected features.

Keywords: Discrete Cosine Transform, Discrete Wavelet Transform, Face Recognition, Feature Selection,
Genetic Algorithm, Particle Swarm Optimization.

1. Introduction

Face recognition (FR) has emerged as one of the most extensively studied research topics
that spans multiple disciplines such as pattern recognition, signal processing and computer
vision. This is due to its numerous important applications in identity authentication, security
access control, intelligent human-computer interaction, and automatic indexing of image and
video databases. Many approaches to face recognitions have been developed; an excellent
survey paper on the different face recognition techniques can be found in [1].
The success of any FR methodology depends heavily on the particular choice of the
features used by the (pattern) classifier. It is known that a good feature extractor for a face
recognition system is claimed to select as more as possible the best discriminant features
which are not sensitive to arbitrary environmental variations such as variations in pose, scale,
illumination, and facial expressions. Feature extraction algorithms mainly fall into two
categories: geometrical features extraction and, statistical (algebraic) features extraction [2 -
8]. The geometrical approach , represent the face in terms of structural measurements and
distinctive facial features that include distances and angles between the most characteristic
International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol. 2, No. 2, June 2009



52

face components such as eyes, nose, mouth or facial templates such as nose length and width,
mouth position, and chin type. These features are used to recognize an unknown face by
matching it to the nearest neighbor in the stored database. Statistical features extraction is
usually driven by algebraic methods such as principal component analysis (PCA), and
independent component analysis (ICA) [6]. These methods find a mapping between the
original feature spaces to a lower dimensional feature space. The shortage of PCA is that it
treats inner-class and out-class equally [3], [4], [5] and therefore it is sensitive to light and
changes of expressions. LDA has higher performance than PCA in face recognition but the
traditional LDA can not provide reliable and robust solution since their separable criterion is
not relevant to classification precision. Alternative algebraic methods are based on transforms
such as downsampling, Fourier transform (FT), discrete cosine transform (DCT), and the
discrete wavelet transform (DWT). Transformation based feature extraction methods such as
the DCT and DWT were found to generate good FR accuracies with very low computational
cost [8]. DCT is one of the approaches used in image compressing which is also used to
extract features [9], [10]. Wavelet analysis has both a good qualities in time domain and
frequency domain which is an ideal tool in analyzing unsteady signals. The DCT and the
DWT feature extraction methods are explained in detail in Section 2.
Feature extraction methods commonly represent the face images with a large set of
features in which features do not contribute equally to the face recognition task. Feature
selection (FS) in pattern recognition involves the derivation of the feature subset from the raw
input data to reduce the amount of data used for classification and simultaneously provide
enhanced discriminatory power. The selection of an appropriate set of features often exploits
the design criteria such as redundancy minimization and decorrelation, and minimization of
the reconstruction error. For many pattern classification problems, a higher number of
features used do not necessarily translate into higher recognition rate [11]. In some cases the
performance of algorithms devoted to speed and predictive accuracy of the data
characterization can even decrease. Therefore, feature selection can serve as a pre-processing
tool of great importance before solving the classification problems. The purpose of the feature
selection is to reduce the maximum number of irrelevant features while maintaining
acceptable classification accuracy. Feature selection is of considerable importance in pattern
classification, data analysis, multimedia information retrieval, biometrics, remote sensing,
computer vision, medical data processing, machine learning, and data mining applications.
The feature selection seeks for the optimal set of d features out of m [11-13]. One possible
approach would be to do an exhaustive search among all possible feature subsets
|
|
.
|

\
|
d
m
and
choose the best one according to the optimization criterion at hand. However, such an
approach is computationally very expensive. Several methods have been previously used to
perform feature selection on training and testing data, branch and bound algorithms [14],
sequential search algorithms [15], mutual information [16], tabu search [17] and greedy
algorithms [12]. In an attempt to avoid the prohibitive complexity FS algorithms usually
involve heuristic or random search strategies. Among the various methods proposed for FS,
population-based optimization algorithms such as Genetic Algorithm (GA)-based method [7],
[18], [19] and Ant Colony Optimization (ACO)-based method have attracted a lot of attention
[20]. These methods attempt to achieve better solutions by using knowledge from previous
iterations with no prior knowledge of features.

In this paper, a face recognition algorithm using a PSO-based feature selection approach is
presented. The algorithm utilizes a novel approach that employs the binary PSO algorithm to
International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol. 2, No. 2, June 2009


53

effectively explore the solution space for the optimal feature subset. The selection algorithm
is applied to feature vectors extracted using the DCT and the DWT. The search heuristics in
PSO is iteratively adjusted guided by a fitness function defined in terms of maximizing class
separation. The proposed algorithm was found to generate excellent recognition results with
less selected features.
The main contribution of this work is:
- Formulation of a new feature selection algorithm for face recognition based on the
binary PSO algorithm. The algorithm is applied to DCT and DWT feature vectors and is used
to search for the optimal feature subset to increase recognition rate and class separation.
- Evaluation of the proposed algorithm using the ORL face database and comparing its
performance with a GA- based feature selection algorithm and various FR algorithms found
in the literature.
The rest of this paper is organized as follows. The DCT and the DWT feature extraction
techniques are described in Section 2. An overview of Particle Swarm Optimization (PSO) is
presented in Section 3. In Section 4 we explain the proposed PSO- based feature selection
algorithm. Finally, Sections 5 and 6 attain the experimental results and conclusion.

2. Feature Extraction

The first step in any face recognition system is the extraction of the feature matrix. A
typical feature extraction algorithm tends to build a computational model through some linear
or nonlinear transform of the data so that the extracted feature is as representative as possible.
In this paper DCT and DWT were used for feature extraction as explained in the following
Sections.

2.1. Discrete Cosine Transform (DCT)

DCT has emerged as a popular transformation technique widely used in signal and image
processing. This is due to its strong energy compaction property: most of the signal
information tends to be concentrated in a few low-frequency components of the DCT. The
use of DCT for feature extraction in FR has been described by several research groups [8],
[9], [10] [21], [22], [23] and [24]. DCT was found to be an effective method that yields high
recognition rates with low computational complexity. DCT exploits inter-pixel redundancies
to render excellent decorrelation for most natural images. After decorrelation each transform
coefficient can be encoded independently without losing compression efficiency. The DCT
helps separate the image into parts (or spectral sub-bands) of differing importance (with
respect to the image's visual quality). DCT transforms the input into a linear combination of
weighted basis functions. These basis functions are the frequency components of the input
data. DCT is similar to the discrete Fourier transform (DFT) in the sense that they transform a
signal or image from the spatial domain to the frequency domain, use sinusoidal base
functions, and exhibit good decorrelation and energy compaction characteristics. The major
difference is that the DCT transform uses simple cosine-based basis functions whereas the
DFT is a complex transform and therefore stipulates that both image magnitude and phase
information be encoded. In addition, studies have shown that DCT provides better energy
compaction than DFT for most natural images.
The general equation for the DCT of an NxM image ) , ( y x f is defined by the following
equation:
International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol. 2, No. 2, June 2009



54

=
+ + =
1
0
1
0
) , ( )] 1 2 (
. 2
.
cos[ )] 1 2 (
. 2
.
cos[ . ) ( ) ( ) , (
N
x
M
y
y x f y
M
u
x
N
u
v u v u F
t t
o o (1)

Where ) , ( y x f is the intensity of the pixel in row x and column y; u = 0,1, ., N-1 and
v=0,1, ., M-1 and the functions ) (u o , ) (v o are defined as

for u,v =0
for u,v=0 (2)

For most images, much of the signal energy lies at low frequencies (corresponding to large
DCT coefficient magnitudes); these are relocated to the upper-left corner of the DCT array.
Conversely, the lower-right values of the DCT array represent higher frequencies, and turn
out to be small enough to be truncated or removed with little visible distortion, especially as u
and v approach the sub-image width and height, respectively. This means that the DCT is an
effective tool that can pack the most effective features of the input image into the fewest
coefficients.
The original face image can be roughly reconstructed only by few DCT coefficients. This
makes the choice of the number of DCT coefficient initially used in the face recognition
system very critical. The effect of the number of DCT coefficients used as features for face
recognition is examined in Section 5. This includes the effect of the number of coefficients on
the quality of the reconstructed image and the recognition rate. The study is extended by
examining the performance of the dynamically generated feature subset generated by the PSO
feature selection algorithm.

2.2. Discrete Wavelet Transform (DWT)

Wavelets have many advantages over other mathematical transforms such as the DFT or
DCT. Functions with discontinuities and functions with sharp spikes usually take substantially
fewer wavelet basis functions than sine-cosine functions to achieve a comparable
approximation. Wavelets have been successfully used in image processing since 1985 [8],
[22], [25], and [26]. Its ability to provide spatial and frequency representations of the image
simultaneously motivates its use for feature extraction. The decomposition of the input data
into several layers of division in space and frequency and allows us to isolate the frequency
components introduced by intrinsic deformations due to expression or extrinsic factors (like
illumination) into certain sub-bands. Wavelet-based methods prune away these variable sub-
bands, and focus on the space/frequency sub-bands that contain the most relevant information
to better represent the data and aid in the classification between different images. There exists
a large selection of wavelet families depending on the choice of the mother wavelet. In this
paper FR using the DWT is based on the facial features extracted from a Haar Wavelet
Transform [8, 22]. The Haar wavelet transform is a widely used technique that has an
established name as a simple and powerful technique for the multi-resolution decomposition
of time series. Earlier studies concluded that information in low spatial frequency bands play a
dominant role in face recognition. In 1986, Sergent [26] shows that the low frequency band
and high frequency band play different roles. The low frequency components contribute to the
global description, while the high frequency components contribute to the finer details
required in the identification task. Sergent has also demonstrated that as human face is a non-

=
N
N
v u
2
1
) ( ), ( o o
International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol. 2, No. 2, June 2009


55

rigid object, it has abundant facial expressions; and expressions influence local spatial
components of face.
The Haar wavelet transform has been proven effective for image analysis and feature
extraction. It represents a signal by localizing it in both time and frequency domains
.Wavelets can be used to improve the image registration accuracy by considering both spatial
and spectral information and by providing multi-resolution representation to avoid loosing
any global or local information. Additional advantages of using the wavelet-decomposed
images include bringing data with different spatial resolution to a common resolution using
the low frequency sub-bands while providing access to edge features using the high frequency
sub-bands.
As shown in Figure 1 at each level of the wavelet decomposition, four new images are
created from the original N x N-pixel image. The size of these new images is reduced to of
the original size, i.e., the new size is N/2 x N/2. The new images are named according to the
filter (low-pass or high-pass), which is applied to the original image in horizontal and vertical
directions. For example, the LH image is a result of applying the low-pass filter in horizontal
direction and high-pass filter in vertical direction. Thus, the four images produced from each
decomposition level are LL, LH, HL, and HH. The LL image is considered a reduced version
of the original as it retains most details. The LH image contains horizontal edge features,
while the HL contains vertical edge features. The HH contains high frequency information
only and is typically noisy and is, therefore, not useful for the registration. In wavelet
decomposition, only the LL image is used to produce the next level of decomposition.



Figure 1. A 3- level wavelet decomposit ion of an N x N- pixel image

Figure 2 shows the decomposition process by applying the 2D Wavelet Transform on a
face image. The original image (shown in Figure (2a)) is decomposed into four sub band
images (shown in Figure (2b)) similarly, 2 levels of the Wavelet decomposition as shown
Figure (2c) can be obtained by applying the wavelet transform on the low frequency band
sequentially.
In Figure (2b), the sub band LL corresponds to the low frequency components in both
vertical and horizontal directions of the original image. Therefore, it is the low frequency sub
band of the original image. The sub band LH corresponds to the low frequency component in
the horizontal direction and high frequency components in vertical direction. Therefore it
holds the vertical edge details. Similar interpretation is made on the sub bands HL and HH.
As the change of facial expressions mainly varies in eyes, mouth and other face muscles,
from the technical point of view, it involves mainly changes of edges. Lets take Figure (2b)
as an example, the horizontal features of eyes and mouth are clearer than its vertical features,
International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol. 2, No. 2, June 2009



56

the sub band HL can therefore depict major facial expression features. The sub band LH, the
vertical features of outline and nose are clearer than its horizontal features, depicts face pose
features. The sub band HH is therefore the most important for rigid object recognition
because it depicts the structure feature of the object. But human faces indeed are non-rigid
objects, the sub band HH is the unstable band in all sub bands because it is easily disturbed by
noises, expressions and poses. Therefore, if wavelet transform is applied to decompose face
images, the sub band LL will be the most stable sub band.


( a) ( b) ( c)
Figure 2. 2D Wavelet decomposit ion of a face image. ( a) The Original I mage ( b)
1- Level Wavelet Decomposit ion ( c) 2- Level Wavelet decomposit ion

3. Particle Swarm Optimization (PSO)

PSO proposed by Dr. Eberhart and Dr. Kennedy in 1995 is a computational paradigm
based on the idea of collaborative behavior and swarming in biological populations inspired
by the social behavior of bird flocking or fish schooling [27], [28], [29], and [30]. Recently
PSO has been applied as an effective optimizer in many domains such as training artificial
neural networks, linear constrained function optimization, wireless network optimization,
data clustering, and many other areas where GA can be applied [29].
Computation in PSO is based on a population (swarm) of processing elements called
particles in which each particle represent a candidate solution. PSO shares many similarities
with evolutionary computation techniques such as GA's. The system is initialized with a
population of random solutions and searches for optima by updating generations. The search
process utilizes a combination of deterministic and probabilistic rules that depend on
information sharing among their population members to enhance their search processes.
However, unlike GA's, PSO has no evolution operators such as crossover and mutation. Each
particle in the search space evolves its candidate solution over time, making use of its
individual memory and knowledge gained by the swarm as a whole. Compared with GAs, the
information sharing mechanism in PSO is considerably different. In GAs, chromosomes share
information with each other, so the whole population moves like one group towards an
optimal area. In PSO, the global best particle found among the swarm is the only information
shared among particles. It is a one-way information sharing mechanism. Computation time in
PSO is significantly less than in GAs because all the particles in PSO tend to converge to the
best solution quickly [29].

3.1. PSO Algorithm
International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol. 2, No. 2, June 2009


57

When PSO is used to solve an optimization problem, a swarm of computational elements,
called particles, is used to explore the solution space for an optimum solution. Each particle
represents a candidate solution and is identified with specific coordinates in the D-
dimensional search space. The position of the i -th particle is represented as X
i
= (x
i1
, x
i2
,..,
x
iD
). The velocity of a particle (rate of the position change between the current position and
the next) is denoted as V
i
= (v
i1
, v
i2
, .., v
iD
). The fitness function is evaluated for each
particle in the swarm and is compared to the fitness of the best previous result for that particle
and to the fitness of the best particle among all particles in the swarm. After finding the two
best values, the particles evolve by updating their velocities and positions according to the
following equations:
) ) *
1 t
i
t
i
t
i
t
i
X X V V + + =
+
best 2 2 i_best 1 1
(g * rand * c (p * rand * c e
(3)
1 t
i
t
i
1 t
i
V X X
+ +
+ = (4)
Where i =(1, 2,. N) and N is the size of the swarm; p
i_best
is the particle best reached
solution and g
best
is the global best solution in the swarm.c
1
and c
2
are cognitive and social
parameters that are bounded between 0 and 2. rand
1
and rand
2
are two random numbers,
with uniform distribution U(0,1). V
max

1 + t
i
V V
max
(V
max
is the maximum velocity). In
equation (3) the first component represents the inertia of pervious velocity. The inertia weight
, is a factor used to control the balance of the search algorithm between exploration and
exploitation; the second component is the "cognitive" component representing the private
experience of the particle itself; the third component is the "social" component, representing
the cooperation among the particles. The recursive steps will go on until we reach the
termination condition (maximum number of iterations K). The pseudo code of the PSO
algorithm is shown in Figure 3.













Figure 3. Pseudo code of t he PSO algorit hm

3.2. Binary PSO and Feature Selection

A binary PSO algorithm has been developed in [30]. In the binary version, the particle
position is coded as a binary string that imitates the chromosome in a genetic algorithm. The
particle velocity function is used as the probability distribution for the position equation. That
is, the particle position in a dimension is randomly generated using that distribution.
The equation that updates the particle position becomes the following:
If rand
3
< 1 t
i
v
e 1
1
+

+
then
1 + t
i
X
=1; else
1 + t
i
X =0 (5)
Initialize parameters
Initialize population
while (number of generations, or the stopping criterion is not met) {
for (i = 1 to number of particles N) {
if the fitness of
t
i
X
is greater than the fitness of
i_best
p

then update
i_best
p
=
t
i
X

if the fitness of
t
i
X
is greater than that of gbest then
then update gbest =
t
i
X

Update velocity vector
Update particle position
Next particle
}
Next generation
}
International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol. 2, No. 2, June 2009



58

A bit value of {1} in any dimension in the position vector indicates that this feature is
selected as a required feature for the next generation, whereas a bit value of {0} indicates that
this feature is not selected as a required feature for the next generation.

4. PSO-Based Feature Selection

The task for the binary PSO algorithm is to search for the most representative feature
subset through the extracted DCT or DWT feature space. Each particle in the algorithm
represents a possible candidate solution (feature subset). Evolution is driven by a fitness
function defined in terms of class separation (scatter index) which gives an indication of the
expected fitness on future trials.

4.1 Chromosome Representation

The initial coding for each particle is randomly produced where each particle is coded to
imitate a chromosome in a genetic algorithm; each particle was coded to a binary alphabetic
string P = F
1
F
2
F
n
, n = 1, 2, , m ; where m is the length of the feature vector extracted by
the DCT or the DWT. Each gene in the m-length chromosome represents the feature
selection, 1 denotes that the corresponding feature is selected, otherwise denotes rejection.
The binary PSO algorithm is used to search the 2
m
genospace for the optimal feature subset
where optimality is defined with respect to class separation. For example, when a 10-
dimensional data set (n=10) P = F
1
F
2
F
3
F
4
F
5
F
6
F
7
F
8
F
9
F
10
is analyzed using binary PSO
to select features, we can select any subset of features smaller than n. i.e. PSO can chose a
random 6 features, F
1
F
2
F
4
F
6
F
8
F
9
by setting bits 1, 2, 4, 6, 8, and 9 in the particle
chromosome. For each particle, the effectiveness of the selected feature subset in retaining the
maximum accuracy in representing the original feature set is evaluated based on its fitness
value.

4.2 Fitness Function

The m-genes in the particle represent the parameters to be iteratively evolved by PSO. In
each generation, each particle (or individual) is evaluated, and a value of goodness or fitness
is returned by a fitness function. This evolution is driven by the fitness function F that
evaluates the quality of evolved particles in terms of their ability to maximize the class
separation term indicated by the scatter index among the different classes [3].
Let
L L
N N N and w w w ,..., , ..., ,
2 1 , 2 1
denote the classes and number of images within each
class, respectively. Let
o L 2 1
M and M ,..., M , M
be the means of corresponding classes
and the grand mean in the feature space, M
i
can be calculated as:

=
= =
i
N
j
i
j
i
i
L i W
N
M
1
) (
. ,..., 2 , 1 ,
1
(6)

Where
, N ,..., 2 , 1 j , W
i
) i (
j
=
represents the sample images from class w
i
and the grand
mean
o
M is:
i
L
1 i
i N
1
0
M N M

=
=
(7)
International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol. 2, No. 2, June 2009


59

Where n is the total number of images for all the classes. Thus, the between class scatter
fitness function F is computed as follows:

=
=
L
i
o i
t
o i
M M M M F
1
) ( ) (
(8)

4.3 PSO-Based Feature Selection Algorithm

An overview of the proposed PSO-based feature selection algorithm is shown in Figure 4.































Figure 4. The PSO- based feat ure select ion algorit hm

4.4. Classifier

A typical and popular Euclidean distance is employed to measure the similarity between
the test vector and the reference vectors in the gallery. Euclidean distance is defined as the
straight-line distance between two points. For N-dimensional space, the Euclidean distance
between two any points p
i
and q
i
is given by:

=
=
N
1 i
2
i i
) q p ( D
(9)
Where p
i
(or q
i
) is the coordinate of p (or q) in dimension i.
Yes
Yes
Generate N particles with random
positions and velocities
Evaluate particles using the fitness
function given in (8)
Return best feature subset
Stopping criteria
satisfied
Yes
P
i best
=
t
i
X
g
best
=
t
i
X

No
No
F(
t
i
X
) >F(P
i_best
)
Update the position of each
particle using(5)
F(
t
i
X
) > F(g
best
)
No
Update the velocity of each
particle using(3)
Initialize parameters c
1
, c
2
, rand
1
, rand
2
, max-iterations, V
max,
International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol. 2, No. 2, June 2009



60


PSO-based
Feature
Selection
Training stage
Recognition stage

Face
Gallery
Training
images

Similarity
Measurement
Test image
Identify the
test image
Feature
Selection
Feature
Matrix
Recognition
Feature
Matrix


DCT


DWT
Feature
Extraction

Feature Extraction
In the application of this approach for face recognition, distances in the feature space from
a query image to every image in the database are calculated. The index of the image which
has the smallest distance with the image under test is considered to be the required index.

5. Experimental Results

The block diagram of the proposed FR system is shown in Figure 5. The block diagram
shows the various steps of processing an input image in the training and recognition stages.
















Figure 5. Block Diagram of t he proposed Face recognit ion syst em.

The performance of the proposed feature selection algorithm is evaluated using the
standard Cambridge ORL gray-scale face database. The ORL database of faces contains a
set of face images taken between April 1992 and April 1994 at the AT&T Laboratories (by
the Oliver Research Laboratory in Cambridge, UK) [20] and [21]. The database is composed
of 400 images corresponding to 40 distinct persons. The original size of each image is
92x112 pixels, with 256 grey levels per pixel. Each subject has 10 different images taken in
various sessions varying the lighting, facial expressions (open/ closed eyes, smiling/ not
smiling) and facial details (glasses/ no glasses). All the images were taken against a dark
homogeneous background with the subjects in an upright, frontal position (with tolerance for
some side movement). Four images per person were used in the training set and the
remaining six images were used for testing. Figure 6 shows sample images from the ORL
face database.
In the experiments carried out in this Section we compare the performance of the
proposed PSO-based feature selection algorithm with the performance of a GA-based
feature selection algorithm. The parameters used for the binary PSO and the GA algorithms
are given in Table 1 and Table 2 respectively.

International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol. 2, No. 2, June 2009


61


Figure 6. Sample images from t he ORL faced dat abase

Table 1. Pso paramet er set t ing Table 2. Ga paramet er set t ing

Swarm size N 30
Cognitive parameters c
1
2
Social parameter c
2
2
Inertia weight 0.6
Number of Iterations 100

For each problem instance, 5 replications are conducted. The average recognition rate is
measured together with the CPU training time and the average number of selected features
for each experiment.

5.1. Experiment 1

In this experiment we test the PSO-based feature selection algorithm with feature vectors
based on various sizes of DCT coefficients. The 2-dimentional DCT is applied to the input
image and only a subset of the DCT coefficients corresponding to the upper left corner of the
DCT array is retained. Subset sizes of 50x50, 40x40, 30x30 and 20x20 of the original
92x112 DCT array are used in this experiment as input to the subsequent feature selection
phase. Figure 7 shows the No. of selected features, training time, and recognition rates for
different feature vector dimensions using the PSO and GA based feature selection
algorithms.
The best average recognition rate of 94.8% is achieved using the DCT (50x50) feature
vector and the PSO-based feature selection algorithm. In this instance, the selection
algorithm reduces the size of the original feature vector by nearly 50%. In general the PSO
and GA selection algorithms have comparable performance in terms of recognition rates but
in all test cases the number of selected features is smaller when using the PSO selection
algorithm. On the other hand, in terms of computational time, GA-based selection algorithm
takes less training time than the PSO-based selection algorithm in all tested instances. Which
indicates that PSO is computationally expensive than GA but the effectiveness' of PSO in
finding the optimal feature subset compared to GA compensates its computational
inefficiency.

The population size 30
Crossover Probability
(PC)
0.
5
Mutation Probability
(PM)
1
Number of Iterations
10
0
International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol. 2, No. 2, June 2009



62



( a)


( b)





( a) ( b) ( c)
Figure 7. Recognit ion result s for different DCT- feat ure based vect ors and Feat ure
select ion Algorit hms. ( a) No. of Select ed Feat ures ( b) Training Time ( c)
Recognit ion Rat e

5.2. Experiment 2

In this experiment the DWT coefficient features have been extracted from each face
image. The 2-dimentional Haar wavelet transform is applied to the input image reducing its
size to 1/4 of its original size. 4-level wavelet decomposition is performed and the
approximation of the input image at each decomposition level is used as a feature vector.
The dimensions of the feature vectors are 46x56, 23x28, 12x14 and 6x8 corresponding to
level-0, level-1, level-2 and level-3 wavelet decompositions respectively.
Figure 8 shows the No. of selected features, training time, and recognition rates for
different feature vector dimensions using the PSO and GA based feature selection
algorithms. The best average recognition rate of 95.2% is achieved using the DWT (12x14)
feature vector and the PSO-based feature selection algorithm using only 88 selected features
and with approximately 35% less selected features than GA. The PSO and GA selection
algorithms have comparable performance in most tested instances but with less selected
features using PSO.











( a) ( b) ( c)

Figure 8. Recognit ion result s for different DWT- feat ure based vect ors and
Feat ure select ion Algorit hms. ( a) No. of Select ed Feat ures ( b) Training Time ( c)
Recognit ion Rat e
International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol. 2, No. 2, June 2009


63

Another DWT-based feature extraction approach was implemented and tested. The 2-
dimentional input image is converted to a 1-dimentional array using raster scan. This is
achieved by processing the image row by row concatenating the consecutive rows into an
array of size 10304. Raster scanning preserves horizontal pixel correlations well but vertical
correlations are lost [31]. The DWT is then applied to the 1-dimentinal image array and used
as feature vector for the feature selection phase. 6-level wavelet decomposition is performed
and the approximation of the input image at each decomposition level is used as a feature
vector. The dimensions of the feature vectors are 2576, 1288, 644, 322, 161, and 81
corresponding to level-1, level-2, level-3, level-4, level-5 and level-6 wavelet
decompositions respectively. Figure 9 shows the No. of selected features, training time, and
recognition rates for different feature vector dimensions using the PSO and GA based feature
selection algorithms.










( a) ( b) ( c)

Figure 9. Recognit ion result s for different DWT- feat ure based vect ors applied t o
1- Diment ional I mage Array and Feat ure Select ion Algorit hms. ( a) No. of
Select ed Feat ures ( b) Training Time ( c) Recognit ion Rat e

The best average recognition rate of 97% is achieved using the DWT(322) feature vector
and the GA-based feature selection algorithm using 262 selected features. This is compared
an average recognition rate of 96.8% with 159 PSO-based selected features for the same test
instance. Experimental results indicate that the recognition accuracy based on features
extracted using the DWT applied to the 1-dimentional raster scan of the input image
outperforms that of the DCT and the DWT of the 2-dimantional input image.
In Table 3, the performance of the proposed algorithm in terms of its recognition rate is
compared to various FR algorithms found in the literature using the ORL database [32].
Table 3 indicates the superiority of the proposed algorithm utilizing the DWT feature
extraction and PSO feature selection. As far as feature selection is concerned the algorithm
selects the optimal number of elements in the feature vector which has a great influence on
the training and recognition times of the algorithm.

6. Conclusion

In this paper, a novel PSO-based feature selection algorithm for FR is proposed. The
algorithm is applied to feature vectors extracted by two feature extraction techniques: the
International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol. 2, No. 2, June 2009



64

DCT and the DWT. The algorithm is utilized to search the feature space for the optimal
feature subset. Evolution is driven by a fitness function defined in terms of class separation.
The classifier performance and the length of selected feature vector were considered for
performance evaluation using the ORL face database. Experimental results show the
superiority of the PSO-based feature selection algorithm in generating excellent recognition
accuracy with the minimal set of selected features. The performance of the proposed
algorithm is compared to the performance of a GA-based feature selection algorithm and
was found to yield comparable recognition results with less number of selected features.

Table 3. Comparison of recognit ion rat es for various FR algorit hms


METHOD

RECOGNITION RATE

TEST CONDITIONS
Hybrid NN: SOM+a
convolution NN

96.2%

DB contained 400 images of 40 individuals.
The classification time less than .5 second for
recognizing one facial image, but training time is
4 hour
Hidden Markov
model
(HMMs)
87%


SVM with a binary
tree

91.21% for SVM and 84.86% for
Nearest Center Classification (NCC)
They compare the SVMs with standard
eigenfaces approach using the NCC

Eigenface 90%

2D-HMM

95%

An average processing time of .22 second for
face pattern with 40 classes.
DCT+PSO FS 94.7 Four images per person were used in the
training set and the remaining six images were
used for testing The training time is less than 3
minutes for all experiments. The average
recognition time for recognizing an input image
is 0.05 sec
DWT +PSO FS 96.8

References

[1] W. Zhao, R. Chellappa, P. J. Phillips, and A. Rosenfeld, Face Recognition: A Literature Survey, ACM
Computing Surveys, vol. 35, no. 4, pp. 399-458, 2003.
[2]R. Brunelli and T. Poggio, Face Recognition: Features versus Templates, IEEE Trans. Pattern Analysis and
Machine Intelligence, vol. 15, no. 10, pp. 1042-1052, 1993.
[3] C. Liu and H. Wechsler, Evolutionary Pursuit and Its Application to Face Recognition, IEEE Trans.
Pattern Analysis and Machine Intelligence, vol. 22, no. 6, pp. 570-582, 2000.
[4] M. A. Turk and A. P. Pentland, Face Recognition using Eigenfaces, Proc. of IEEE Conference on
Computer Vision and Pattern Recognition, pp. 586-591, June 1991.
[5] L. Du, Z. Jia, and L. Xue, Human Face Recognition Based on Principal Component Analysis and Particle
Swarm Optimization -BP Neural Network, Proc 3
rd
Conference. on Natural Computation (ICNC
2007),vol. 3, pp. 287-291, August 2007.
[6] X. Yi-qiong, L. Bi-cheng and W. Bo, Face Recognition by Fast Independent Component Analysis and
Genetic Algorithm, Proc. of the 4
th
International Conference on Computer and Information Technology
(CIT04), pp. 194-198, Sept. 2004.
[7] X. Fan and B. Verma, Face recognition: a new feature selection and classification technique, Proc. 7
th

Asia-Pacific Conference on Complex Systems, December 2004.
[8] A. S. Samra, S. E. Gad Allah, R. M. Ibrahim, Face Recognition Using Wavelet Transform, Fast Fourier
Transform and Discrete Cosine Transform, Proc. 46
th
IEEE International Midwest Symp. Circuits and
Systems (MWSCAS'03), vol. 1, pp. 272- 275, 2003.
International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol. 2, No. 2, June 2009


65

[9] C. Podilchuk and X. Zhang, Face Recognition Using DCT-Based Feature Vectors, Proc. IEEE
International Conference on Acoustics, Speech and Signal Processing (ICASSP96), vol. 4, pp. 2144-2147,
May 1996.
[10] Z. Yankun and L. Chongqing, Efficient Face Recognition Method based on DCT and LDA, Journal of
Systems Engineering and Electronics, vol. 15, no. 2, pp. 211-216, 2004.
[11] C.-J. Tu, L.-Y. Chuang, J.-Y. Chang, and C.-H. Yang, Feature Selection using PSO-SVM, International
Journal of Computer Science (IAENG), vol. 33, no. 1, IJCS_33_1_18.
[12] E. Kokiopoulou and P. Frossard, Classification-Specific Feature Sampling for Face Recognition, Proc.
IEEE 8
th
Workshop on Multimedia Signal Processing, pp. 20-23, 2006
[13] A. Y. Yang, J. Wright,Y. Ma, and S. S. Sastry, Feature Selection in Face Recognition: A Sparse
Representation Perspective, submitted for publication, 2007.
[14] P. M. Narendra and K. Fukunage, A Branch and Bound Algorithm for Feature Subset Selection, IEEE
Trans. Computers, vol. 6, no. 9, pp.917-922, Sept. 1977.
[15] P. Pudil, J. Novovicova, and J. Kittler, Floating Search Methods in Feature Selection, Pattern Recognition
Letters, vol. 15, pp. 1119-1125, 1994.
[16] B. Roberto, Using Mutual Information for Selecting Features in Supervised Neural Net Learning, IEEE
Trans. Neural Networks, vol. 5, no. 4, pp. 537-550, 1994.
[17] H. Zhang and G. Sun, Feature Selection Using Tabu Search Method, Pattern Recognition Letters, vol. 35,
pp. 701-711, 2002.
[18] D.-S. Kim, I.-J. Jeon, S.-Y. Lee, P.-K. Rhee, and D.-J. Chung, Embedded Face Recognition based on Fast
Genetic Algorithm for Intelligent Digital Photography, IEEE Trans. Consumer Electronics, vol. 52, no. 3,
pp. 726-734, August 2006.
[19] M. L. Raymer, W. F. Punch, E. D. Goodman, L.A. Kuhn, and A. K Jain, Dimensionality Reduction Using
Genetic Algorithms, IEEE Trans. Evolutionary Computation, vol. 4, no. 2, pp. 164-171, July 2000.
[20] H. R. Kanan, K. Faez, and M. Hosseinzadeh, Face Recognition System Using Ant Colony Optimization-
Based Selected Features, Proc. IEEE Symp. Computational Intelligence in Security and Defense
Applications (CISDA 2007), pp 57-62, April 2007.
[21] F. M. Matos, L. V. Batista, and J. Poel, Face Recognition Using DCT Coefficients Selection, Proc. of the
2008 ACM Symposium on Applied Computing, (SAC08),pp. 1753-1757, March 2008.
[22] M. Yu, G. Yan, and Q.-W. Zhu, New Face recognition Method Based on DWT/DCT Combined Feature
Selection, Proc. 5
th
International Conference on Machine Learning and Cybernetics, pp. 3233-3236,
August 2006.
[23] Z. Pan and H. Bolouri, High Speed Face Recognition Based on Discrete Cosine Transform and Neural
Networks, Technical Report, Science and Technology Research Center (STRC), University of
Hertfordshire.
[24] Z. M. Hafed and M. D. Levine, Face Recognition Using Discrete Cosine Transform, International
Journal of Computer Vision, vol. 43, no. 3, pp. 167-188. 2001.
[25] D.-Q. Dai and H. Yan, Wavelets and Face Recognition, in Face Recognition, K. Delac and M. Grgic, Eds.
I-Tech, Vienna, Austria, 2007, pp.558.
[26] J. Sergent, Microgenesis of face perception,. In: H.D. Ellis, M.A. Jeeves, F. Newcombe and A. Young,
Editors, Aspects of Face Processing, Nijhoff, Dordrecht (1986).
[27] J. Kennedy and R. Eberhart, Particle swarm optimization, Proc. IEEE International Conference on
Neural Networks, pp. 1942-1948, 1995.
[28] R. C. Eberhart and J. Kennedy, A New Optimizer Using Particles Swarm Theory, Proc. Roc. 6
th

International Symp. Micro Machine and Human Science, pp. 39-43, Oct. 1995.
[29] R. C. Eberhart and Y. Shi, Comparison between Genetic Algorithms and Particle Swarm Optimization,
Proc. 7
th
international Conference on Evolutionary Programming, pp. 611-616, 1998.
[30] J. Kennedy and R. C. Eberhart, A Discrete Binary Version of the Particle Swarm Algorithm, Proc. IEEE
International Conference on Systems, Man, and Cybernetics, vol. 5, pp. 4104-4108, Oct. 1997.
[31] J. Modayilt, H. Chengt, and X. Lii, Experiments in Simple One-Dimensional Lossy Image Compression
Schemes, Proc. 1997 International Conference on Multimedia Computing and Systems (ICMCS '97), pp.
614-615, 1997.
[32] A. S. Tolba, A.H. El-Baz, and A.A. El-Harby Face Recognition: A Literature Review, International
Journal of Signal Processing, vol. 2, no. 2, pp. 88-103. 2006



International Journal of Signal Processing, Image Processing and Pattern Recognition
Vol. 2, No. 2, June 2009



66