International Journal of Computer Trends and Technology (IJCTT) volume 4 Issue 8 August 2013

ISSN: 2231-2803 http://www.ijcttjournal.org Page 2556

Face Recognition Using Multi-Support Vector Machines
A.Swathi
1
, Dr.R.Pugazendi
2

Department of Computer Science,
K.S.Rangasamy College of Arts and Science, Tiruchengode, Tamilnadu, India



Abstract- In Face Identification task, an image of an
unknown person is matched to a gallery of known people.In
Face Verification task, is to accept or deny the identity
claimed by a person.Therefore, given two face images, the
goal is to decide whether two images are same or not.In
previous research PLS method which used for scalable to the
gallery size and modify one-against-all approach to use a
treebased structure. At each internal node of the tree, a
binary classifier based on PLS (Partial Least Square)
regression is used to guide the search of the matching subject
in the gallery. The use of this structure provides substantial
reduction in the number of comparisons when a probe sample
is matched against the gallery. There are very accurate
techniques to perform face identification in controlled
environments, particularly when large numbers of samples
are available for each face. As an enhancement this paper
proposes to use SVM (Support Vector Machine) classifier to
recognize the person. An experimental result shows that the
recognition accuracy and time efficiency has been increased
by using a SVM.

Keywords Binary Classification, Discriminant
Hyperplanes, Support Vector Machine, Partial Least Square.

I. INTRODUCTION

Face Recognition is an important part of the reported
security incidents, biometrics-based techniques are
becoming a very interesting authentication and identification
method for both IT Security providers and actually, for an
increasing number of sectors of the civil society.Biometrics
identifiers are built from a unique, physical or behavioral
trail of an individual for automatically recognizing or
verifying the identity [18]. As such, they provide a solution
to the problem of unequivocal identification of users, and,
hence, can efficiently prevent identity theft and
unauthorized access attacks.The research on face recognition
has recently proposed a new method[2] [3] based on one
class Support Vector Machines[17], which consider very
promising and worth looking at into more detail. Support
Vector Machines, the learning approach originally
developed by Vapnik et al[19], represent a powerful pattern
recognition method, able to deal with sample sizes of order
of hundred of thousands instances [10]. They have been
used till now for solving several practical problems, like
isolated handwritten digit recognition [14], speaker
identification [13], face detection [11] and text categorization
[9].

Mainly, Support Vector Machines are used for multi-
class classification [5], in which any new object is assigned
to one of a predefined set of classes [4].The proposed
face recognition procedure is built on the work presented in
[3], but also introduces a new component based approach, in
which the one-class SVM algorithm is applied to the main
components of the human face (eyes, nose and mouth). The
combined global and local approach performs better that the
original algorithm. The paper also proposes a different
feature extraction method instead of raw gray level features
and presents the experimental results obtained with three
different light normalization procedures.
The paper is organized as follows: the next section
presents the basic operational flow of the recognition
procedure and outlines the novelty of the proposed approach
with respect to the prior work. In the third section the
experimental results are presented. The paper ends with a
conclusions and future work section.

II.RELATED WORK

Partial Least Squares (PLS) has been widely adopted
as the most promising face recognition algorithm. PLS is
used extensively in all forms of analysis from neuroscience
to computer graphics because it is a simple, non-
parametric method of extracting relevant information from
confusing data sets. With minimal additional effort PLS
provides a roadmap to reduce a complex data set to a lower
dimension to reveal sometimes hidden, simplified structure
that often underlie it.
PLS is an orthogonal transformation of the coordinate
system in which the pixels are described. The main idea of
the principal component analysis is to find the vectors which
best describe the distribution of face images within the
entire image space and it aims to extract a subspace where
the variance is maximized. PLS is performed by projecting
a new image into the subspace called face space spanned by
the eigenfaces and then classifying the face by comparing its
position in face space with the positions of known
individuals.
A face image in 2-dimension with size N N can also
be considered as one dimensional vector of dimension
N
2
. The main idea of the principle component is to find the
vectors that best account for the distribution of face images
within the entire image space. These vectors define the
subspace of face images, which is call face space. Each
of these vectors is of length N
2
, describes an N N image,
International Journal of Computer Trends and Technology (IJCTT) volume 4 Issue 8 August 2013
ISSN: 2231-2803 http://www.ijcttjournal.org Page 2557

and is a linear combination of the original face images.
These vectors are the eigenvectors of thecovariance matrix
corresponding to the original face images, and because they
are face like in appearance, they are referred as
eigenfaces.The below algorithm is named as PLS which is
used for classification.

Step 1: The normalized training image in the N-
dimensional space is stored in a vector of size N. Let the
normalized training face image set,

T={X
1
, X
2.
X
N
} where X={x
1
, x
2
... x
m
}
T

Step 2: Each of the normalized training face images are
mean centered. This is done by subtracting the mean face
image fromeach of the normalized training images. The
mean image is represented as a column vector where each
scalar is the mean of all corresponding pixels of the training
images,
X X Xt
t
=
Where the average of the training face image set is defined as
1
1
N
i
X Xn
N
=
=


Step 3: Once the training face images are centered, the
next process is to create the Eigenspace which is the reduced
vectors of the mean normalized training face images.
The training images are combined into a data matrix of size
N by P, where P is the number of training images and each
column is a single image
2
1
{ , ,......, } P X X X X =
Step 4: The column vectors are combined into a data
matrix which is multiplied by its transpose to create a
covariance matrix. The covariance is defined as:
T
X X O=
Step 5: The Eigen values and corresponding eigenvectors
are computed for the covariance matrix using J acobian
transformation,

V V O = A
Where v is the set of eigenvectors associated with the Eigen
values .

Step 6: Order the eigenvectors V
i
V according to their
corresponding eigenvalues
i
from high to low with non-
zero eigenvalues. This matrix of eigenvectors is the
eigenspace zero eigenvalues. This matrix of eigenvectors is
the eigenspace.
{ }
1 2
, ,....,
i P
V V V V =

A. Problem Identification
Need to handle the problem of insufficient training
data.
Need to achieve high performance when only a
single sample per subject is available.
Need to reduce the number of comparisons made for
verification.

IV. PROPOSED METHOD

Support vector machines are supervised learning models
with associated learning algorithms that analyze data and
recognize patterns.It used for classification and regression
analysis. SVM takes a set of input data and predicts, for each
given input, which of two possible classes forms the output,
making it non-probabilistic binary linear classifier.
SVM maps input vectors to a higher dimensional vector
space where an optimal hyper plane is constructed.Among
the many hyper planes available, there is only one hyper
plane that maximizes the distance between itself and the
nearest data vectors of each category.This hyper plane which
maximizes the margin is called the optimal separating hyper
plane and the margin is defined as the sum of distances of the
hyper plane to the closest training vectors of each category.
Alternatively, an SVM-based face detector can be
designed. This work, extend the idea of [6, 7] to the multi-
view situation based on pose estimation. It is interesting to
notice that:
1. While the eigenface method models the probability density
of face patterns, the SVM-based method only models the
boundary between faces and non-faces;
2. By solving a quadratic programming problem, the SVM-
based method is guaranteed to converge to the global
optimum;
3. The solution is expressed directly by a subset of
important training examples called Support Vectors.
B. Linear case
Consider a set of l vectors {xi}, xi R
n
, 1 i l,
representing input samples and set of labels {yi }, yi {1},
that divide input samples into two classes, positive and
negative. If the two classes are linearly separable, there
existsa separating hyperplane (w, b) defining the function, in
equation (1)
f (x) =< w x > +b, (1)
And sgn (f (x)) shows on which side of the hyperplane x
rests,in other words the class of x. Vector w of the
separating hyperplane can be expressed as a linear
combination of xi (often called a dual representation of w)
with weights i:
i i
l i
i
y

s s
=
1
(2)
The dual representation of the decision function f (x) is then:

s s
+ > < =
l i
i i i
b x x y x f
1
. ) ( (3)
Training a linear SVM means finding the embedding strengths
{i} and offset b such that hyperplane (w, b) separates pos-
International Journal of Computer Trends and Technology (IJCTT) volume 4 Issue 8 August 2013
ISSN: 2231-2803 http://www.ijcttjournal.org Page 2558

itive samples from negatives ones with a maximal margin.
Notice that not all input vectors {xi} might be used in the
dual representation of w; those vectors xi that have weight
i > 0 and form w are called support vectors.

C. Non-linear case
In real-life problems it is rarely the case that positive
and negative samples are linearly separable. Non-linear
support vector classifiers map input space X into a feature
space F via a usually non-linear map : X F, x 1
(x) and solve the linear separation problem in the feature
space by finding weights i of the dual expression of the
separating hyperplanes vector w:

s s
=
l i
i i i
x y
1
) ( (4)
While the decision function f (x) takes the form

s s
+ < =
l i
i i i
b x x y x f
1
) ( ). ( ) ( (5)
Usually F is
function K(x, y) that computes thedot product in F , K(x,
y) =< (x) (y) >. The decision function (5) can then be
computed by just using the kernel function and it can also
be shown that finding the maximum margin separating
hyperplane is equivalent to solving the following
optimization problem
(

s s l j i
j i j i j i i
x x K y y
, 1
) , (
2
1
max

(6)

s s
= s s s s
l i
i i i
y l i c
1
0 , 1 , 0

Where positive C is a parameter showing the trade-off
between margin maximization and training error
minimization.The kernel function K avoid working directly
in feature space F. After solving (6), offset b can be chosen
so that the margins between the hyperplane and
(

+ =

s s l i
i i i
b x x k y x f
1
) , ( sgn )) ( sgn(
(7)
The two classes of sample images are equal then have
our decision function Commonly used kernels include
polynomial kernels K(x, y) =(x +y) d and the Gaussian
kernel K(x, y) =exp (||xy||2).This implementation use the
Gaussian kernel, however one of the interesting points for
further research is approaches for choosing an optimal kernel
for the given input data.
V. EXPERIMENTAL RESULTS
In order to demonstrate the effectiveness of SVMs for
object recognition under varying illumination conditions, it
conducted a number of experiments by using the Face
Database. It is known that face recognition under varying
illumination conditions is a very difficult task because
variations due to changes in illumination are usually larger
than those due to changes in face identity [13]. The
performance of generative methods for face recognition [8,
1] implies that the assumption that there is no intersection
among illumination cones is approximately satisfied.












Fig 1.Components-based approach detection and training
Componenet based approach detection and training
presented in figure 1.It represents the features.
D. Face image database
The database consists of images of 10 individuals in 9
poses acquired under 64 different point light sources and an
ambient light: 5850 images in total. The coordinates of the
left eye, right eye, and mouth are appended for images in the
frontal pose, and the coordinate of the face center is
appended for images in other poses. Each image is assigned
to one of 5 subsets according to the angle between the
direction of the light source and the optical axis of a camera.


Fig.2.Illumination images in database for single person 16
illumination Images are there.
Fig.2.shows the different illumination conditions of image
database.


Fig.3.Input Image
Fig.4.Enhanced Image
International Journal of Computer Trends and Technology (IJCTT) volume 4 Issue 8 August 2013
ISSN: 2231-2803 http://www.ijcttjournal.org Page 2559

Image Enhancem
Fig 3 and 4 represents input image and enhanced image.
After Image Enhancement, feature extractions are performed
such as Hog, lbp, colour features and Gabor features.








Fig.5.Non overlapping image left side eye








Fig.6.Non overlapping image right side eye





2.Non overlbing image rideside eye.

Fig.7.Non overlapping image left side Nose&Mouth








Fig.8.Non overlapping image right side Nose&mouth
Non overlapping image right and left side eye represent in
figure 5 and 6 and Non overlapping image left and right side
nose and mouth represent in figure 7 and 8.It display the
hog,lbp,color,gabor features.

The table 1 represents the classification using SVM for
identifying the person.

TABLE 1: CLASSIFICATION USING SVM


TABLE 2: CLASSIFICATION USING PLS

Table 2 represents the classification using PLS for identifying
the person.

0
5
10
15
20
Person 1Person 2Person 3Person 4Person 5
I
l
l
u
m
i
n
a
t
i
o
n
Persons


Person 1 Person 2 Person 3 Person 4 Person 5
Person 1 17 0 0 0 0
Person 2 0 17 0 0 0
Person 3 0 0 17 0 0
Person 4 0 0 0 17 0
Person 5 0 0 0 0 17

Person
1
Person
2
Person
3
Person
4
Person
5
Person
1
17 0 1 0 0
Person
2
0 17 0 0 0
Person
3
0 0 15 0 0
Person
4
0 0 0 17 0
Person
5
0 0 1 0 17

International Journal of Computer Trends and Technology (IJCTT) volume 4 Issue 8 August 2013
ISSN: 2231-2803 http://www.ijcttjournal.org Page 2560

Fig.9. SVM classification
0
5
10
15
20
Person 1Person 2Person 3Person 4Person 5
I
l
l
u
m
i
n
a
t
i
o
n
Persons

Fig.10.PLS Classification
The classifications chart of SVM and PLS is represented in
figure 9 and 10.It analyse the accuracy performance of SVM
is better than PLS.

97
98
99
PLS SVM
accuracy

Fig.11.Comparison with PLS and SVM
Fig.11 shows the accuracy of SVM and PLS.The
experimental results shows SVM classification is higher than
PLS classification and accuracy of SVM classification is
98.7%.

While comparing SVM with PLS, The table1 shows the
classification of persons. There is no misclassification in
SVM on person identification 3.Table 2 shows there is
misclassification on person identification 3 in PLS, so SVM
has highest accuracy then PLS.

VI. CONCLUSIONS AND FUTURE WORK

This paper discussed the problem of object recognition
under varying illumination conditions, in this paper, an
efficient Support vector machine approach to face
recognition based on wavelet transform is proposed. The
exsisting result shows that the face recognition performance
is relatively unaffected even though there is a transformation
of the face including translation, small rotation and
illumination.The SVM algorithm is used, because it is the
basic and straight forward method for feature extraction. It
provides efficient results and requires less storage.
In the present study, the experiment is conducted by
using face images. However, the proposed method should be
applicable to non-Lambertian objects when illumination
cones of objects are approximated by low dimensional
subspaces.Therefore, the future work plan is to confirm the
effectiveness of SVM method for objects with various
reflectance properties.The future efforts will be on the
recognition of face images in dynamic video sequences and
real time tasks using fuzzy logic.

REFERENCES

[1] R. Basri and D. J acobs, Lambertian reflectance and linear subspaces,
In Proc. IEEE ICCV 2001, pp.383390, 2001.

[2] P. Belhumeur, J . Hespanha, and D. Kriegman, Eigenfaces vs.
Fisherfaces: recognition using class specific linear projection, IEEE
Trans.PAMI, 19(7), pp.711720, 1997.

[3] P. Belhumeur and D. Kriegman, What is theset of images of an object
under all possible lighting conditions?, Intl. J . Computer Vision,
28(3), pp.245260, 1998.
[4] K. Bennett and E. Bredensteiner, Duality and geometry in SVM
classifiers, In Proc. Intl. Conf. Machine Learning (ICML 2000),
pp.6572, 2000.
[5] R. Brunelli and T. Poggio, Face recognition: featuresversus templates,
IEEE Trans. PAMI, 15(10), pp.10421052, 1993.
[6] H. Chen, P. Belhumeur, and D. J acobs, In search of illumination
invariants, In Proc. IEEE CVPR 2000, pp.254261, 2000.
[7] R. Duda, P. Hart, and D. Stork, Pattern Classification, J ohn Wiley
&Sons, New York, 2001.
[8] A. Georghiades, P. Belhumeur, and D. Kriegman, Fromfew to many
illuminations cone models for face recognition under variable lighting
and pose, IEEE Trans. PAMI, 23(6), pp.643660, 2001.
[9] G. Guo, S. Li, and K. Chan, Face recognition by support vector
machines, In Proc. IEEE FG 2000, pp.195201, 2000.
[10] P. Hallinan, A low-dimensional representation of human faces for
arbitrary lighting conditions, In Proc. IEEE CVPR 94, pp.995999,
1994.
[11] B. Heisele, P. Ho, and T. Poggio, Face recognition with support vector
machines: global versus component-based approach, In Proc. IEEE
ICCV 2001, pp.688694, 2001.
[12] Y. Moses, Y. Adini, and S. Ullman, Face recognition: the problemof
compensating for changes in illumination direction, In Proc. ECCV
94, pp.286296, 1994.
[13] H. Murase and S. Nayar, Visual learning and recognition of 3-D
objects fromappearance, Intl. J . Computer Vision, 14(1), pp.524,
1995..
[14] P. Phillips, Support vector machines applied to face recognition,
Advances in Neural Information Processing Systems 11, pp.803809,
1998.
[15] M. Pontil and A. Verri, Support vector machines for 3D object
recognition, IEEE Trans. PAMI, 20(6), pp.637646, 1998.
[16] R. Ramamoorthi and P. Hanrahan, On the relationship between
radiance and irradiance: determining the illumination fromimagesof a
convex Lambertian object, J . Opt. Soc. Am. A, 18(10), pp.2448
2459,2001.
[17] R. Ramamoorthi and P. Hanrahan, A signal-processing framework for
inverse rendering, In Proc. ACM SIGGRAPH 2001, pp.117128,
2001.