Using Gabor Filters Features for Multi-Pose Face Recognition in Color Images

Zhi-Kai Huang, Wei-Zhong Zhang, Hui-Ming Huang, Ling-Ying Hou

Department of Machinery and Dynamic Engineering,
Nanchang Institute of Technology, Nanchang, Jiangxi 330099, China
huangzhik2001@yahoo.com.cn
Abstract
Human face recognition plays an important role in
applications such as video surveillance, human computer
interface, and face image database management. This
paper presents an improved face recognition method for
Multi-Pose Face Recognition in color images, which
addresses the problems of illumination and pose variation.
At first, color Multi-Pose faces image features were
extracted based on Gabor wavelet with different
orientations and scales filters, then the mean and
standard deviation of the filtering image output are
computed as features for face recognition. In addition,
these features were fed up into support vector machine
(SVM) for face recognition. Experimental results show
that successful face recognition over a wide range of
facial variations in color, position, scale, orientation, 3D
pose, and expression in images from Stereo-pair database.
1. Introduction
Human face recognition plays an important role in
applications such as video surveillance, human computer
interface, and face image database management. In recent
years face recognition has received substantial attention,
but still remained very challenging in real applications.
Despite the variety of approaches and tools studied, face
recognition is not accurate or robust enough to be used in
uncontrolled environments.
There are two main approaches for face recognition,
holistic and geometric [1]. Geometric approaches
dominated in the 1980s where simple measurements such
as the distance between the eyes and shapes of lines
connecting facial features [2] were used to recognize
faces, while holistic methods became very popular in the
1990s with the well known approach of Eigenfaces [3].
Even though holistic methods such as neural networks are
more complex to implement than their geometric
counterparts, their application is much more straight
forward, whereby an entire image segment can be reduced
to a few key values for comparison with other stored key
values and no exact measures or knowledge such as eye
locations or the presence of moustaches needs to be
known. The problem with this grab all approach was
that noise, occlusions such as glasses and any other non-
face image attribute could be learned by the holistic
algorithm and become part of the recognition result even
though such factors are not unique to faces [5].
Color representation is a problem of significant
importance in the fields of computer vision and image
processing. Color-based image classification is inherently
invariant to transformations in the image that are due to
rescaling, translation, distortion, and rotation. So, many
color spaces have been suggested for computer vision
tasks. They attempt to obtain a better representation of the
color. A 3-dimensional color space is defined as a
mathematical representation of a set of colors, where each
color is mapped to three coordinates. Commonly used for
image include RGB, YIQ, L*u*v*, HSV (or HSL, HSB),
and the opponent [4].
In this paper, we first convert the original RGB image
to the different color spaces image. Multi-pose face
representation and based on the output of Gabor filters in
different color space is computed. Finally, the obtained
Gabor feature vectors are fed up into support vector
machine for face recognition and the results of face
recognition in different color space have been compared.
2. Color Model and Face features extracted
using Gabor Filters
2.1 Color Model
A combination of real world physical characteristics
determines what the human vision system perceives as
color. A color space is a mathematical representation of
these characteristics. Color Spaces are always three-
dimensional [7].
There are many possible color space definitions:
Digital imagery often uses the red/green/blue color
space, known simply as RGB. The cyan/yellow/magenta
space, known as CYM, is used in printing. Hue, saturation,
Second International Symposium on Intelligent Information Technology Application
978-0-7695-3497-8/08 $25.00 2008 IEEE
DOI 10.1109/IITA.2008.29
399
Second International Symposium on Intelligent Information Technology Application
978-0-7695-3497-8/08 $25.00 2008 IEEE
DOI 10.1109/IITA.2008.29
399
Second International Symposium on Intelligent Information Technology Application
978-0-7695-3497-8/08 $25.00 2008 IEEE
DOI 10.1109/IITA.2008.29
399
and intensity, (or HSI) is the color space typically used by
artists. Intensity-chromaticity color spaces, YUV and YIQ,
are used for television broadcast.
A particular Gabor elementary function can be used as
the mother wavelet to generate a whole family of Gabor
wavelets.
( , ) g x y RGB is an m-by-n-by-3 image array whose three
planes contain the red, green, and blue components for the
image. . It is based on fact that almost all the colors can be
expressed by mixing in the fixed proportions of three
selected clusters of light of the properly chosen width of
spectrum.
A two-dimensional Gabor function and its
Fourier transform can be written as:
( , ) G u v
jWx
y x
y x g
y x y x
t
o o o to
2 (
2
1
exp
2
1
) , (
2
2
2
2
+
(
(

+
|
|
.
|

\
|
=
(1)
In 1931 the CIE(Commission Internationale de
l'Eclairage)developed the XYZ color system, also called
the "normcolour system". This system is often represented
as a two-dimensional graphic, which more or less
corresponds to the shape of a sail. The red components of
a color are tallied along the horizontal axis of the
coordinate plane and the green components along the
vertical axis. In this way every colorcan be assigned a
particular point on the coordinate plane. The spectral
purity of colors decreases as you move left along the
coordinate plane. What is not taken into consideration in
this model is brightness.
)
`

=
2
2
2
2
) (
[
2
1
exp ) , (
v u
v W u
v u G
o o
(2)
1
2
u
x
o
to
=
1
2
v
v
o
to
=
1 j =
Where: , ,and .
Gabor functions form a complete but nonorthogonal
basis set. Expanding a signal using this basis provides a
localized frequency description. Let be the
mother Gabor wavelet, then this self-similar filter
dictionary can be obtained by appropriate dilations and
rotations of through the generation function:
( , ) g x y
In the NTSC (National Television System Committee)
color space, image data consists of three components:
luminance (Y), hue (I), and saturation (Q). The first
component (luminance) represents grayscale information,
while the last two components make up chrominance
(color information). One of the main advantages of this
format is that grayscale information is separated from
color data, so the same signal can be used for both color
and black and white sets.
( , ) g x y
, (3)
) , ( ) , ( y x g a y x g
m
mn
' ' =

eger n m a int , , 1 = >

where
) sin cos (
) sin cos (
u u
u u
y x a x
x y x a y
m
m
+ = '
' + = '

(4)
k
nt
u =
The YCbCr color space is widely used for digital video.
In this format, luminance information is stored as a single
component (Y), and chrominance information is stored as
two color-difference components (Cb and Cr). Cb
represents the difference between the blue component and
a reference value. Cr represents the difference between
the red component and a reference value.
(5)
where and indicate the orientation and scale of
the filter respectively.
k
is the total number of
orientations and
n
is the total number of scales in the
filter bank. The scale factor is meant to ensure that
the energy is independent of [4].
n k
m
a

The HSV color space (hue, saturation, value) is often

used by people who are selecting colors (e.g., of paints or
inks) from a color wheel or palette, because it corresponds
better to how people experience color than the RGB color
space does.
m
( , ) I x y
For a given image , the decomposed image at
scale
k
and orientation is :
n
1 1
)
1
,
1
( ) , ( ) , ( dy dx y y x x
mn
g y x I y x
mn
W
}
-
=
2.2 Face features extracted using Gabor Filters
(6)
where- indicates the complex conjugate.
One of the crucial aspects of texture classification is
the extraction of proper and representative textural
features. Here, we consider the features generated by
Gabor filters [4]. Basically, Gabor filters are a group of
wavelets, with each wavelet capturing energy at a specific
frequency and a specific direction. Expanding a signal
using this basis provides a localized frequency description,
therefore capturing local features/energy of the signal.
Multi-Pose faces image features can then be extracted
from this group of energy distributions.
It is assumed that the local texture regions are spatially
homogeneous, and the mean
mn

and the standard

deviation
mn
o
of the magnitude of the transform
coefficients are used to represent the region for
classification purposes:
dxdy y x W
mn mn mn }}
=
2
) ) , ( ( o dxdy y x W
mn mn }}
= ) , (
and
400 400 400
A feature vector is now constructed using
mn

and
mn
o
as feature components.
The texture descriptor for
k
scales and orientations
is given by:
n
|
nk nk
features | o o o , , , , ,
11 11 00 00
=
While the size of the filter bank is application
dependent, experimentation has shown that a bank of
filters tuned to combinations of four scales, and six
orientations, at 60-degree intervals, is sufficient for the
analysis of bark image. The architecture of the bark
classification system would look something like the one
presented in Figure 1.
(7)
R(H,Y,etc)
(Gabor filter)
features(1...48)
face image
Fig. 1. Overall architecture of the face recognition system
After the Gabor features of Multi-Pose faces image
have been extracted which had described in section 2, the
second task is that recognition of color face image using
support vector machine (SVM).
3 Image Data and Experimental Results
In this experiment, we used the Multi-Pose faces image
database from [6], which addresses the problems of
illumination and pose variation. It consists of stereo pairs
of 70 subjects (35 males, 35 females), recorded from 6
different viewpoints. We selected 160 pictures of 20 faces
in 8 poses capture from former database. In the training
step it splits the training dataset into two non-overlapping
parts, 60% of it for training and 40% of it for validation.
These images were recorded at a resolution of 480 x 640
pixels, with a bit depth of 16 bits/pixel. Some color Multi-
Pose faces images are shown in Fig.2.
Fig. 2. Example of multi-pose faces image images, The first row shows the images from the left camera of the stereo-pair, the
second row shows the images taken from the right camera.
The Gabor texture feature is based on the scheme
proposed in [6], which uses a bank of Gabor filters to
extract image features information at multiple scales and
orientations. In our experiment, we use six scales and four
orientations, which result in a feature vector of 48. The
image features were calculated using dedicated
subroutines written in Matlab 7.0 language for a 1G
memory, Pentium 2.4GHz PC computer.
We used the quantity average recognition rate
defined as below to compare the results.
G(S,I ,etc)
(Gabor filter)
features(49...96)
B(V, Q,etc)
Gabor filter
features(97...144)
Color space
transform
SVM
classifier
Result of
recognition
401 401 401
Average Recognition Rate = (Number of face
classified truly) (Total number of classified face image)
%
The average recognition rates have been presented in
Figure 3.
?
+
b
8
80
8?
8+
8b
88
90
9?
9+
YC|C |S\ \TSC 0u)
^vu_ ko_u`t`ou kut
!
Fig. 3. Average Recognition Rate in Different Color Space
In order to compare average recognition with gray
image, the corresponding results are shown in Figure
3.From the classification performances shown in Figure 3,
we found that: 1) for different color space, there exists an
optimal choice which achieves the best recognition rate.
We observed that in YIQ color space can achieves the
better performance in bark recognition experiments. 2) In
contrast to only use gray information in image, increased
color channel information can improve average
recognition rate.3) As for the spatial sampling resolutions,
adopting the more Gabor image feature that can improve
accuracy of bark classification, but it will lead to a time-
consuming computation.
4. Conclusion
This paper proposes a multi-Pose faces image
recognition algorithm, in which Gabor feature
representation in different color space is extracted and
SVM classifier are employed. After extracting Multi-Pose
faces image features, Gabor filter the image has been
filtered with four orientations and six scales filters, and
then the mean and standard deviation of the image output
are computed. Finally, the obtained Gabor feature vectors
are fed up into support vector machine (SVM) for
classification. This paper also demonstrates that when a
bank of Gabor filters is applied to an image, there are
strong relationships between the outputs of the different
filters. These relationships are used to devise a new
feature which is capable of describing Multi-Pose faces
image information in a concise manner. The performance
of color space features is found to be better than that of
the features which just extracted from gray image. The
experimental results show this approach can be used to
automatically identify the plant categories. The future
study will focus on how to extract more efficient features
from bark images to improve the classification accuracy
further.
References
[1] V. Bruce, P.J.B. Hancock and A.M. Burton (1998)
Comparisons between human and computer recognition of
faces, Proceedings Third IEEE International Conference
on Automatic Face and Gesture Recognition, 14(16),
pp.408-413,
[2] G. Yongsheng and M.K.H. Leung. (2002) Face
recognition using line edge map, IEEE Transactions on
Pattern Analysis and Machine Intelligence, 24(6), 764-779.
[3] M. Turk and A. Pentland. (1991) Eigenfaces for
recognition. Journal of Cognitive Neuroscience, 3, pp.71-
86
[4] Zhi-Kai Huang, Hui-Ming Huang, Yun-Min,Xie, Ling-
Ying Hou. (2008) Face Recognition Using Model of the
Scene Shape Unsupervised Image Segmentation Using EM
Algorithm by Histogram, Journal of Computational
Information Systems: Vol. 4,pp.919-928
[5] J.E. Meng, W. Chen and W. Shiqian. (2005) High speed
face recognition based on discrete cosine transform and
RBF neural networks, IEEE Transactions on Neural
Networks, 16(3),pp.679 691
[6] R. Fransens, C. Strecha, L. Van Gool.(2005) Parametric
Stereo for Multi-Pose Face Recognition and 3D-Face
Modeling, ICCV 2005 workshop "Analysis and
Modeling of Faces and Gestures"
http://cvlab.epfl.ch/data/stereoface/
[7] Zhi-Kai Huang and Zhi-Feng Wang. (2007) Bark
Classification Using RBPNN in Different Color Space,
Neural Information Processing Letters and Reviews Vol.
11, No. 1, pp.7-13
[8] PRTools Classification Toolbox,
http://www.prtools.org/cgi-bin/twiki.bin/view
402 402 402