2004 International Conference on Image Processing (ICIP)

A SVM-BASED METHOD FOR FACE RECOGNITION USING A WAVELET-

PCA REPRESENTATION OF FACES
Majid Safari, Mehrtash T. Harandi and Babak N. Araabi
Control and Intelligcnt Processing Centcr of Exccllcncc
Dcpaitment of E l d i i c a l and Computer Engineering, Univcrsitv of Tehran, Tehran, Iran
&
School of cognitive science (SCS)
lnstitutc for studies in theoretical physics and mathematics (IPM), Tehran, Iran
111 .wtfbri:&e. .slmrif..edfr , i i i h ( ? r ~ i i i [ i ~ ~U!.~ , cireK-
~ UL'. . < 7 1 ~ ~ 1 ( i h i ~ $nc.
l I l .ir-

ABSTRACT faccs and introduced the Eigcnlhce method Ibr l i c e

T1iI.s paper propose.^ a rzew nietliod oJ Jace
repmwrrtatim n,/iich is irsedfo~face ~ ~ ~ p i t iby o SVh
t i I.
For.fnce rqneserltatiori we /raw irseci a iwo-step ndroii,
Di.ccrute Wavelet Trarisfor7n (111V77
Jrrr t~vo-~iiirierisiorral
is irscri io transfor711 the faces to a iiiore riiscririririoteci
.space arid theii Priricipal Cornporieirt Aria/,wis (PCA) is
applied The proposed rr~ethorlproduced a .sigrirJcarzt
inrpiowrrieirt which incliicies a sirbstarifial reduction iri
ewer rate and in tirite ofprocessing cltririg the obtaining
PCA or?honor?rtalbasis.
1. INTRODUCTION
Several features (height, iris, fingqrint, voice, face)
have been uscd for human identification. Face recognition
is a natura!, user friendly and non-contact way for human
identiBcation/authentication.Face recognizers are used in
the r e n t years for identity authentication, access control,
surveillance, human-computer interface and marl
environments [!I.
Current Face recognizers can be categorized into two
goups, component-based face recopizers and image-
based face recognizers [?]. Component-based systems use
only a few features that are extracted from image. These
features are the important and mostly invariable parts of
face (e.g. eyes, nose, mouth and chin). However, this
method has limited applications because of its difficult
implementation and its unreliability in some cases. Image-
based systems use a pixel intensity matrix that yields a
high dimensional feature space. The image-based method
is more desirable for researchers due to its ease of
implementation and robustness. In order to build an
accurate and fast face recognizer, a lower dimension
representation from the raw image must be obtained.
Principal Components Analysis (PCA), Linear
Discriminant Analysis (LDA) and Independent
Component Analysis (ICA) have been used for this
purpose. Turk successfully used the PCA to represent
recognition [4]. PCA delivers a low dimensional
rcprcscntation of an image and cxtincts a compact glohal
structure h-om it. On the other hand, IJM extracts
discriminaton. ~caturcs so that the class- can be
differentiated in an optimum way. 'Thcsc two lincnr
methods can he changed to some nonlincai- methods by
means of Kernel functions (i.e. KPCA and KDA, [7]).
Applying PCA on wavelet diagonal detail coefficients of
imagcs was examincd by Yucla [9]. After represcnting
faces in a lower dimension space, a classifier such as
nearest distance classifier, k-nearest neighborhood, neural
network or support vector machine (SVM) [3], [8])can be
applied to perform the recognition process. Recently,
interests on using SVM have been increased because of its
promising s p e d and accuracy.
In this work S V M is used as a face classifier with
RBF and polynomial kernel function. We also utilized
DWT as a preprocessing tool for face representation. The
new proposed method decreases the error rate by more
than 2.5% comparing to the PCA. The proposed method
also reduces processing time for determining Eigenfaces
significantly.
In the next section we will lntrcduce our face
representation system and proposed face recognition
method. In section 3 the simulation results are reported
and comparisons to Eigenface method are made. Finally
conclusions are given in section 4.
2. METHODOLOGY
The first step of every pattern recognition system is
feature extraction. The extracted features must he low in
dimension and optimum in the discriminatov power
simultaneously in order to decrease the time and error of
recognition. To achieve these goals a two-stage approach
was used for extracting face features. First the mean
image is subtracted from all the training images in order to
normalize them. A suhband representation of normalized
data was obtained by means of DWT. By applying PCA

0-7803-8554-3/04/$20.00 02004 IEEE. 853

on the suhband images Eigenhces arc obtained (Fig. I).
PCA orthonoimal bases in order to project images into a
lower dimension space for recokmition. These projected
vectors are sent to SVM system to train with database
vectors or recognize thc rcccived image identity.
In our work the various f o m s of wnvelet function are
used in an attempt to find the optimum one. Thc two-
dimensional intensity matrix of the input image w o k s as
the input of DWT system and the output is made by
saving thc approximation coefficients and discarding the
three sets of detail coexffcienls. ‘Thc approximation
coeflicients aic thc input oI‘ thc PCA system to detenninc
cigcnlaccs (Fig. I ) in the leaming phase or to project
images tu coordinates that are defined hy eigcnfaccs in thc
recognition phase (Fig. 2 ) .
3.1. DWT-PCA
Two-dimensional wavelet transform is used in this
method. Thc recognition rate is dependent to the selected
wavelet function and the level of decomposition. In order
to obtain the hest result, these parameters were examined
by experiments. Fig. 3 shows emi- rate I‘or difftrent levels
of decomposition in the training phase. The hest results
are ohtained when two level of decomposition is used. In
Fig. 3 and Fig. 5 , DWT-PCA and PCA methods arc
compared with dilkrcnt orders of polynomial SVM and
KBF S V M respectively.

1 le* mMet - REF Sm mfh t e s t Gamma

-.-
-
_b_

2 lejel waxlo(- Pdynmial Sun Order 1

2 iejel wade Pdynmial Sun Older 2
~

-.-
4
-
2 lelel wadet REF h m v i l h k i t Gamma
~

3 lejel ydMd - Pdynmial Sun Order 1

3 leiel rad8 - Pdvnmlal sun order 2

0
Fig 1 Blmk diagram for ohtaining Eigcnfaces by D W - P C A prooess. 50 150 XI)

Fig 3 E m r mte YS. number of Eigenfaces for one (light gray lints), two
(block lines), and three (dark gray lines) levels of DWT decomposition.

Fig 2 Block diagram ofprojecting images into eigcnlvces space.

3. EXPERIMENTAL RESULTS

For evaluating our new method, we have used the

ORL’ database. The ORL face database (developed at the
Olivetti Research Laboratory, Cambridge, U.K.) is
composed of 400 images with ten different images for
each of the 40 distinct subjects. The variations of the
images are across pose, size, time, and facial expression. Fig 4 Error rate VQ. different number of Eigenfacea for PCA method
All the images were taken against a dark homogeneous (gray lines) and DWT-PCAmethod (black lines) using polynomial Svhrl
classitisrs with ditfPrent orders.
background with the subjects in an upright, frontal
position, with tolerance for some tilting and rotation of up
to about 20”. There is mme variation in scale of up to
about 10%. The spatial and grey-level resolutions of the
images are 9’2 112 and 256, respectively. Five images
from each individual are selected randomly for training
and the five rest images are used for testing.

Olivstti Research Laboratory

854
 - m.M , BFSVMGmmail

- PCA-DWT , REF SVM Gmma=O 1

RBF SYM Ganma=O 1

Fig 5 - Emorrole VI. dilkrsnt numbcrot'EigcnCaces for PCA (gray
liner) ;md llic DWT-IC4 mcthods (hlack lincs) using RBF SVhf
clnssilicr with 1uo dift'efcrenl viilues uf(;smma.
With respect to these ligui-cs, it is concludcd that the
ncw method decreases the error ratc by more than 3%
relative to thc I'CA method and maintains the property of
having the silme or better discriminatoiy power with less
features. A great advantage is that the time of processing
for determining Eigcdaces decreased sipificantly. This
is due to thc fact that wavclet transform reduces the size
of images which must he processed to obtain eigenfaces.
This cffect significantly overcomes the effect of extra
time-consuming pmcesing which is done hy wavelct
transform.
The improvement of new DWT-PCA method @y
two-dimensional wavelet transform) inspired M to
examine some related methods to discover their effects.
Fig 6 - Enor ralc vs. diffcrcnl numhcr oCEigsnhces for on<_llircc and 5
I w d s ofO\VT dccomporilion (one dimcnsional w v e l r l ) .
3.3. DWT without PCA
In this expcrimcnt, the DWT approximation
coefficients arc 1.eordered hy zigzag scanning and the
SVM inpul wctors are constlucted directly ft-om the
rearranged approximation coefiicients. Here face
recognition was camicd done without using PCA.
In Fig. 7 error ratc for three levels of wavelet
decomposition for different number of prescwed
approximation coeffcient is shown One can ohtain
similar mor rate when using a large numher of
coefficients and dimension reduction ability doesn't exist.
In the case of four level of wavelet decomposition, the
number of coefficients can he rduced significantly but
error rate is more than the PCA method (Fig. 8).

3.2. D W - P C A Using One-Dimensional Wavelet

Transform

In this experiment one-dimensional wavelet transform

is used instcad of two-dimensional one. The results show
,
that using onc-level of decomposition makes no
4 \
significant changes compared to PCA method. One can
see by increasing levels of decomposition discriminatoiy
power and efficiency reduces extremely (Fig. 6).
This result can he interpreted in this way, the
correlation between each pixel and its' neighbors in the
same row is neglected. In fact this transform supposes
that the image space is a onedimensional space which
reduces between pixels information and consequently the
efficiency. M 60 10 Po 90 rm 110 1 8 ,w 140 If0

Fig 7 Errorrate YS. different number ofcoefficients forthree

level ofDWT deoompasition.

855
 [J] M. A. Turk and A. P. Pentl:ind. "Eigenfaces for.
recognition." Jur~molo/ Cognilive Ne,,roseimce, Vol. 3 .
1991, pp. 71-86
[51 K . Jonsson. J. Matas. J. Kinlerand Y . Li, "Lramingsupport
vectors for facc veriticatian and recognition." Pruc. I E E
bilenrrlior,al Co,&rmce on iluloamlic Foce and Crstrtrc
Rceoprilion, 2000. pp. 208-2 13.
[6] S. 0 Mallnt, "A Thcory lor Multiresolution Signal
Decoiuposition The Wavelct Reprcsentntion."
IEEE.Tro,isoc/iorrs OII Padem .4,mlvsi.~ mid dfachinc
l ~ m l l i ~ e Vol.
m ~ ~II.. 19x9. pp. 674-693.
[7] M . H. Yang, "Kcmcl EigenFaces YS Kcmel Fishcrfaczs:
Face Recognition using Kemcl Methods," Proc. o//lte F$/z
b z l . Cot$ on ~ 4 ~ ~ l o Face~ ~ ~r ~
a d~ Ge.ylwe
l i c Recogrriliorr.
May 2002. pp. 215-220.
[8] V. Vapnik. "Statistical Lcaming T h c o ~ . "John Wilcy &
Sons. New York, 1998.
[9] P. C. Yucln. I).Q.I h i and G.C. Tong, "WnvelGt-boscdPCA
for JUIL:~I facc recognition.''Proc. I
('ullfi.l~<~rrce ion, I!l9X.
ut1 Illloge AllO!l:Fir O l d Ilrll~r.p~'~t
Fig 8 Emor mlc vs. diffemnt numhcr olcorllicicntr for four IWEI pp. 223.228.
of mvr dccvmpoJitim.

3.4. DCT-PCA

Discrete Cosine Transfoim (DCT) is used instead of

DWT in the first stage. l h e results are similar to PCA and
it seems that the DCT has no effect on images. The
representation vectors in two methods (PCA and DCT-
PCA) are also similar in most of coefficients and the few
differences are neglected hy SVM.

4. CONCOLUSIONS
In this paper a new method tor face rcpreszntation and
recognition using DWT and SVM classiIier is proposed.
DWT is used as a preprocessing tool which improves the
recognition performance significantly. This improvement
includes a substantial reduction in mor rate and
processing time of obtaining PCA orthonormal hasis
representation. These improvements lead an enhancement
in efficiency or, in other words, security of system and
moving toward real-time face recognizer. It is noteworthy
that Lhese improvements didn't affect other suitable
properties of PCA method adversely.

5. REFRENCES

R. Chellnppa, C. L. Wilson, and S. Sirohey, "Human and

machine recognition of faces: A survey,'' Proc. IEEE, May
1995,pp. 705-741.
R. Brunelli, and T. Poggio, '"Face recognition: Feahttures
versus templates," IEEE Transaclions on Pallern Anatvsis
andiliadine Iiihlligmce, Vol. 15, n. IO, 1993, pp. 1042-
1052.
C. Cortes, and V. Vapnik, "Support-Vector Nehvorks,"
Machine Learning, Vol. 20, n. 3, Sept. 1995, pp. 273-297.

856