See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/258651934

Face Recognition using Eigenvector and Principle Component Analysis

Article in International Journal of Computer Applications · July 2012

DOI: 10.5120/7811-0947

CITATIONS READS

11 7,266

3 authors, including:

Sanjit Kumar Saha Md. Al-Amin Bhuiyan

Jahangirnagar University King Faisal University
14 PUBLICATIONS 61 CITATIONS 76 PUBLICATIONS 818 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Md. Al-Amin Bhuiyan on 03 February 2017.

The user has requested enhancement of the downloaded file.

 International Journal of Computer Applications (0975 – 8887)
Volume 50 – No.10, July 2012

Face Recognition using Eigenvector and Principle

Component Analysis

Dulal Chakraborty Sanjit Kumar Saha Md. Al-Amin Bhuiyan

Lecturer Lecturer Professor
Department of Information & Department of Computer Science Department of Computer Science
Communication Technology, & Engineering, & Engineering, Jahangirnagar
Comilla University, Comilla, Comilla University, Comilla, University, Savar, Dhaka-1342,
Bangladesh Bangladesh Bangladesh

ABSTRACT The third group [14, 15] uses feature vectors extracted from a
Face recognition is an important and challenging field in frontal view of the face. Although there are three different
computer vision. This research present a system that is able to approaches to the face recognition problem, there are two
recognize a person’s face by comparing facial structure to that basic methods from which these three different approaches
of a known person which is achieved by using frontal view arise.
facing photographs of individuals to render a two-dimensional The first method is based on the information theory concepts,
representation of a human head. Various symmetrization which is also known as the principal component analysis
techniques are used for preprocessing the image in order to method. Based on the Karhunen-Loeve expansion in pattern
handle bad illumination and face alignment problem. We used recognition, M. Kirby and L. Sirovich [6, 7] have shown that
Eigenface approach for face recognition. Eigenfaces are any particular face could be economically represented in
eigenvectors of covariance matrix, representing given image terms of a best coordinate system that they termed
space. Any new face image can then be represented as a linear "eigenfaces". Later, M. Turk and A. Pentland [16] have
combination of these Eigenfaces. This makes it easier to proposed a face recognition method based on the eigenfaces
match any two given images and thus face recognition approach.
process. The implemented eigenface-based technique The second method is based on extracting feature vectors
classified the faces 95% correctly. from the basic parts of a face such as eyes, nose, mouth, and
chin. In this method, with the help of deformable templates
Keywords and extensive mathematics, key information from the basic
Principle component analysis, eigenvector, eigenvalue, parts of a face is gathered and then converted into a feature
eigenface, faces recognition. vector. L. Yullie and S. Cohen [17] played a great role in
adapting deformable templates to contour extraction of face
images.
1. INTRODUCTION After three decades of research effort, the Eigenface approach
The face is primary focus of attention in social life which merged as the first real successful demonstration of automatic
plays a major role in conveying identity and emotion. human face recognition. This is one of the methods which can
Although face recognition is challenging, the human ability to be classified as appearance-based methods that use the whole
recognize faces is remarkable. We can recognize thousands of face region as the raw input to a recognition system. The goal
faces learned throughout our lifetime and identify familiar of an appearance-based face recognition algorithm is
faces at a glance even after years of separation. This skill is essentially to create low-dimensional representations of face
quite robust, despite large changes in the visual stimulus due images to perform recognition. In contrast, geometric feature-
to viewing conditions, expression, aging, and distractions such based methods attempt to distinguish between faces by
as glasses, beards or changes in hair style. comparing properties and relations between facial features,
Face recognition has fundamental importance in our social such as eyes, mouth, nose and chin. As a consequence,
relationship, being extremely important for our simple success of these methods depends heavily on the feature
everyday activities. It is a very high level task and has many extraction and measurement process.
applications. Developing a computational model of face The low-dimensional representation of faces in the Eigenface
recognition is quite difficult, because faces are complex, approach is derived by applying Principle Component
multi-dimensional visual stimuli. A formal method of Analysis (PCA) to a representative dataset of images of faces.
classifying faces was first proposed by Francis Galton [1, 2] The system functions by projecting face images onto a feature
in 1888. During the 1980’s work on face recognition remained space that spans the significant variations among known face
largely dormant. Since the 1990’s, the research interest in face images. These significant features are termed “Eigenfaces”
recognition has grown significantly. because they are the principal components of the set of
Investigations by numerous researchers [3, 4, 5] over the past training face images .These features do not necessarily
several years have indicated that certain facial characteristics correspond to facial features such as eyes, nose and ears. They
are used by human beings to identifying faces. There are three merely capture the image points that cause meaningful
major research groups which propose three different variations between the faces in the database that allow them to
approaches to the face recognition problem. The largest group be differentiated. Face images are then classified within the
[6, 7, 8] has dealt with facial characteristics which are used by low-dimensional model using a nearest-neighbor classifier.
human beings in recognizing individual faces. The second The Eigenface approach works well on test images unaffected
group [9, 10, 11, 12, 13] performs human face identification by illumination changes. It is a well-known fact that
based on feature vectors extracted from profile silhouettes. intrapersonal differences (e.g. illumination effects, poses)

42
 International Journal of Computer Applications (0975 – 8887)
Volume 50 – No.10, July 2012

cause more variations between face images than interpersonal recognition pre-requisite such as gray image conversion and
differences (identity). To handle this variability, methods scaling into a suitable sized image.
usually take one of two approaches: measure some property in
the image that is invariant or at least insensitive to 3.1.1 Conversion to Gray Image and Scaling
illumination effects, or model the object in order to predict the Detected face is converted to grayscale using equation (1) and
variations caused by changes in illumination. Solutions that scaled to 60  60 pixel using equation (2) and saved as a gray
follow the former approach are so far still elusive, and may jpg image. Linear interpolation technique was employed to
never exist at all. This suggests that appearance-based determine the scaled output image.
methods, which derive low-dimensional models of the face Ri  G i  Bi
images used for training, are the only answer to this Gr i  , i=1,…..., M  N (1)
3
challenging problem.
th
Where Gr i is the gray level value of i pixel of the gray
2. EIGENVECTORS AND image. Ri , Gi , Bi corresponds to red, green, blue value of
EIGENVALUES th
“Eigen” is a German word meaning “proper” or “own”. An
the i pixel in the color image.

  p
 y q 
eigenvector of a matrix is a vector such that, if multiplied with p
Q x q , y  P x x q ,
q
the matrix, the result is always an integer multiple of that y  (2)
vector. This integer value is the corresponding eigenvalue of  60 60
 
the eigenvector. This relationship can be described by the
Where, we want to re-scale image
equation M × u = c × u, where u is an eigenvector of the
P[(0,0)  ( x p , y )] to image Q0,0  132,132
p
matrix M and c is the corresponding eigenvalue.
Type of image Red Green Blue
2.1 Eigenfaces
Eigenfaces are the set of eigenvectors which used in computer
vision problem for human face recognition. They can be
simply defined as the eigenvectors which represent one of the
dimension of face image space. All eigenvectors have an RGB
eigenvalue associated to it and the eigenvectors with the
largest eigenvalues provide more information on the face
variation than the ones with smaller eigenvalues.
Grayscale
2.2 Principle Component Analysis (PCA) Figure 1. Conversion of RGB image to Grayscale image
Technique
PCA is one of the most successful techniques that have been 4 FACE RECOGNITION
used in face recognition. The objective of the Principal A simple approach to extracting the information contained in
Component Analysis is to take the total variation on the an image of a face is to somehow capture the variation in a
training set of faces and to represent this variation with just collection of face images, independent of any judgment of
some little variables. When we are working with great features, and use this information to encode and compare
amounts of images, reduction of space dimension is very individual face images.
important. PCA intends to reduce the dimension of a group or In mathematical terms, the principal components of the
to space it better so that the new base describes the typical distribution of faces, or the eigenvectors of the covariance
model of the group. The maximum number of principal matrix of the set of face images, treating an image as point (or
components is the number of variables in the original space. vector) in a very high dimensional space is sought. Each
Even so to reduce the dimension, some principal components image location contributes more or less to each eigenvector,
should be omitted. so that it is possible to display these eigenvectors as a sort of
ghostly face image which is called an "eigenface".
3. FACE DETECTION Sample face images and the corresponding eigenfaces are
To locate the face, an image pyramid is formed from the shown in Figure 2 and in Figure 3 respectively. Eigenfaces
original image. An image pyramid is a set of copies of the can be viewed as a sort of map of the variations between
original image at different scale, thus representing a set of faces.
different resolutions. A mask is moved pixel wise over each
image in the pyramid and at each position, the image section
under the mask is passed to a function that assesses the
similarity of the image section to a face. If the similarity value
is high enough, the presence of a face at that position and
resolution is assumed. From the position of the face, a first
estimate of the eye position can be derived. A search for the
exact eye position is started. The positions yielding the
highest similarity values are taken as final estimates of the eye
positions.

3.1 Face Image Normalization

After the face area has been detected, it is normalized before
passing to the face recognition module. We apply a sequence
of image pre-processing techniques so that the image is light Figure 2. Training set face images
and noise invariant. We also need to apply some standard face Each individual face can be represented exactly in terms of a
linear combination of the eigenfaces. Each face can also be

43
 International Journal of Computer Applications (0975 – 8887)
Volume 50 – No.10, July 2012

approximated using only the "best" eigenfaces, those that have vector form. Let these images be 1 ; 2 ;........M .The
the largest eigenvalues, and which therefore account for the
average face of the set is
most variance within the set of face images. The best M M
1
eigenfaces span an M-dimensional subspace which we call the
"face space" of all possible images.
  i (3)
M i 1
Each face image differs from the average face of
the distribution, and this distance is calculated by subtracting
the average face from each face image. This gives us new
image space .
 i  i  ( i = 1,2,………..M ) (4)

4.2 Calculating Eigenfaces Definitions:

An N  N matrix A is said to have an eigenvector X, and
corresponding eigenvalue λ if
AX  X (5)
Evidently, Eq. (7) can hold only if
det A  I  0 (6)
which, if expanded out, is an Nth degree polynomial in λ
Figure 3. Eigenfaces with highest eigen values, that whose root are the eigenvalues.
were calculated from the sample training set, given in A matrix is called symmetric if it is equal to its transpose,
A A aij  a ji
Figure 2 T
or (7)

Kirby and Sirovich [6, 7] developed a technique for efficiently It is termed orthogonal if its transpose equals its inverse,
AA  A A  I
T T
representing pictures of faces using principal component (8)
analysis. Starting with an ensemble of original face images, Finally, a real matrix is called normal if it commutes with is
they calculated a best coordinate system for image transpose,
compression, where each coordinate is actually an image that
AA  A A
T T
they termed an "eigenpicture". In this research, we have (9)
followed the method which was proposed by M. Turk and A.
Theorem: Eigenvalues of a real symmetric matrix are all
Pentland [16] inorder to develop a face recognition system
real. Contrariwise, the eigenvalues of a real non symmetric
based on the eigenfaces approach. They argued that, if a
matrix may include real values, but may also include pairs of
multitude of face images can be reconstructed by weighted
complex conjugate values. The eigenvalues of a normal
sum of a small collection of characteristic features or
matrix with non degenerate eigenvalues are complete and
eigenpictures, perhaps an efficient way to learn and recognize
orthogonal, spanning the N dimensional vector space. Let the
faces would be to build up the characteristic features by
experience over time and recognize particular faces by training set of face images be 1 ; 2 ;........M then the
comparing the feature weights needed to approximately average of the set is defined by
reconstruct them with the weights associated with known 1 M
individuals.   i (10)
The basic steps involved in Face Recognition using
M i 1
Eigenfaces Approach are as follows: Each face differs from the average by the vector
1. Acquire initial set of face images known as Training Set  i  i  (11)
(  i ). An example training set is shown in Figure 2. This set of very
2. The average matrix  has to be calculated. Then subtract large vectors is then subject to principal component analysis,
this mean from the original faces (  i ) to calculate the image which seeks a set of M ortho-normal vectors, U n which best
vector (  i ). describes the distribution of the data. The k
th
vector, U k is
1 M chosen such that
  i
M i 1
1 M
 k  M  uTk  n  2
(12)
 i  i  n 1
is a maximum, subject to
3. Find the covariance matrix C by
u l u k   lk 
T
M 1, if l = k (13)
1
 n 
T
c n
 T
AA 0, otherwise
M n 1
The vectors uk and scalars  k are the eigenvectors and
4. Compute the eigenvectors and eigenvalues of C.
5. The M’significant eigenvectors are chosen as those with the eigenvalues, respectively of the covariance matrix
largest corresponding eigenvalues 1 M
 n  n  AAT
T
6. Project all the face images into these eigenvectors and form c (14)
M n 1
the feature vectors of each face image.
where the matrix A  [ 1 ,  2 ,......., M ] . The covariance
4.1 Training Set of Images
matrix C, however is N  N
2 2
Let the training set consists of M images representing M real symmetric matrix, and
image classes. Each of these images can be represented in

44
 International Journal of Computer Applications (0975 – 8887)
Volume 50 – No.10, July 2012

determining the N 2 eigenvectors and eigenvalues is an faces are not in the database. For known images, we found
that the mean difference is almost zero. So the training set of
intractable task for typical image sizes. images easily recognize with the corresponding image in the
If the number of data points in the image space is less than the database. For unknown images, we found an exponential
2
dimension of the space (M < N , there will be only M-1, mean difference between database images and training set of
images that mean the input images are not recognized with
rather than N 2 meaningful eigenvectors. The remaining any of database images. The resultant analysis has given
eigenvectors will have associated eigenvalues of zero. below with table and diagram.
T
Consider the eigenvectors vi of A A such that Table 1. Training procedure of image
A A V l   l vl
T
(15)
Training Procedure
Pre multiplying both sides by A, we have
No. of images taken for 35  3
AA A V l   l A vl
T
(16) the training procedure
Size 60  60
from which we see that A vi are the eigen vectors of Format JPG
C  AAT . (17) Output Normalized images
Following these analysis, we construct the M  M matrix of the extracted face
images
L  AT A (18) Eigenface
Lmn    n and find the M eigenvectors, vl , of L.
T Average Eigenvector
where m Mean Image
These vectors determine linear combinations of the M training
set face images to form the eigenfaces u l . Input image database in day light condition:
M
ul   vlk  k , l = 1,2,…….,M (19)
k l
With this analysis, the calculations are greatly reduced, from
the order of the number of pixels in the images N   to the
2

order of the number of images in the training set (M).

5. IMPLEMENTATION AND RESULT

ANALYSIS
The experiment is implemented over a training set of 105 (3
each of 35 Images). Each image is in RGB level which is
normalized to gray level and has dimension of 60  60. There
are 7 subjects in the training set. Each subject has 5 images
with frontal view with  50 different poses (like left, right, up
and down). In our face database which images we have taken,
they are all the students of department of Computer Science &
Engineering, Jahangirnagar University of Bangladesh. Their
average age is 22 years. All images are the same size. No
image contains sunglass, beard or mustaches in face area. Generated eigen faces of image database in day light
Face images are taken in sunny bright light which we called
condition:
day light condition and in dark light or low light condition.
Then we use histogram equalized condition for face
recognition. So we can say that in recognition procedure we
use different illumination condition and all this condition it
can recognize the known or unknown face.
Firstly we construct overall average image. This is the image
which is formed by adding all images and dividing by number
of images in training set. The eigenvectors of covariance
matrix is formed by combining all deviation of training set’s
image from average image.
After finding overall average image, we have to find the
eigenvectors of the covariance matrix. Since there are 35
images in the training set we need to find 35 eigenvectors
that are used to represent our training set.
After finding eigenvectors, we use it to recognize the the
training set of images. In our experiment we consider the
training set of images for three conditions (day light, low light
and histogram equalized) for both known and unknown faces
where known faces are stored at face database and unknown

45
 International Journal of Computer Applications (0975 – 8887)
Volume 50 – No.10, July 2012

Generated eigen vector for day light condition image Result of Recognition Procedure in day light condition:
database:

Recognition procedure in day light condition

Total number of images taken for the test 10

Number of recognized images 10

Number of misrecognized images Nil

(no store in databases)
Number of misrecognized images Nil
(store in databases)
Accuracy 100%

Input image database in low light condition:

Result of known image in day light condition:

Image
Input position Mean Resultant Result
Image in difference Position
database

1 0.0021 1 TRUE

9 0.0048 9 TRUE

20 0.0027 20 TRUE

29 0.0039 29 TRUE

34 0.0021 34 TRUE Generated eigen faces for image database in low light
condition:
Result of unknown image in day light condition:

Image
position Mean Resultant Result
Input in difference Position
Image database
1.0e+005
NULL * 0.3705 3 TRUE

1.0e+005
NULL * 0.2870 30 TRUE

1.0e+005
NULL * 0.5358 20 TRUE

1.0e+005
NULL * 1.2142 12 TRUE
1.0e+005
NULL * 0.1076 4 TRUE

46
 International Journal of Computer Applications (0975 – 8887)
Volume 50 – No.10, July 2012

Generated eigen vector for low light condition: Result of recognition procedure in low light condition:

Recognition procedure in low light condition

Total number of images taken for the test 10

Number of recognized images 10

Number of misrecognized images Nil

(no store in databases)

Number of misrecognized images Nil

(store in databases)

Accuracy 100
%

Input image database in histogram equalized condition:

Result of known image in low light condition:

Image
Input position Mean Resultant Result
Image in difference Position
database

22 0.0018 22 TRUE

26 0.0017 26 TRUE

1 0.0038 1 TRUE

9 0.0019 9 TRUE

15 0.0020 15 TRUE

Result of unknown image in low light condition:

Generated eigen faces in histogram equalized condition:
Image
Input position Mean Resultant Result
Image in difference Position
database
1.0e+005
NULL * 0.4373 1 TRUE

1.0e+005
NULL * 0.1483 20 TRUE

1.0e+005
NULL * 0.0997 32 TRUE

1.0e+005
NULL * 1.3318 21 TRUE
1.0e+005
NULL * 0.4225 22 TRUE

47
 International Journal of Computer Applications (0975 – 8887)
Volume 50 – No.10, July 2012

Generated eigen vector for histogram equalized image Result of recognition procedure in histogram equalized
database: condition:

Recognition procedure in day light condition

Total number of images taken for the test 10

Number of recognized images 10

Number of misrecognized images Nil

(no store in databases)

Number of misrecognized images Nil

(store in databases)

Accuracy 100%

6. CONCLUSION
In this paper eigenface based face recognition has been
described. The eigenface approach for face recognition
Result of known image in histogram equalized condition: process is fast and simple which works well under constrained
environment. It is one of the best practical solutions for the
Image Mean Resultant Result problem of face recognition. Eigenfaces method is a principal
Input position difference Position component analysis approach, where the eigenvectors of the
Image in covariance matrix of a small set of characteristic pictures are
database sought. These eigenvectors are called eigenfaces due to their
resemblance of face images. Recognition is performed by
28 0.0045 28 TRUE obtaining feature vectors from the eigenvectors space.

Many applications which require face recognition do not

1 0.0004 1 TRUE require perfect identification but just low error rate. So instead
of searching large database of faces, it is better to give small
set of likely matches. By using Eigenface approach, this small
18 0.0031 18 TRUE set of likely matches for given images can be easily obtained.
For given set of images, due to high dimensionality of images,
the space spanned is very large. But in reality, all these
15 0.0028 15 TRUE images are closely related and actually span a lower
dimensional space. By using eigenface approach, we try to
reduce this dimensionality. The lower the dimensionality of
22 0.0035 22 TRUE this image space, the easier it would be for face recognition.
Any new image can be expressed as linear combination of
these eigenfaces. This makes it easier to match any two
Result of unknown image in histogram equalized
images and thus face recognition.
condition:
Input Image Mean Resultant Result One of the limitation for eigenface approach is in the
Image position difference Position treatment of face images with varied facial expressions and
in with glasses. Also as images may have different illumination
database conditions. This can be removed by RMS (root mean square)
1.0e+005 contrast stretching and histogram equalization.
NULL * 1.1427 31 TRUE
In the present work, we used 105 face images as face
1.0e+005 database. The procedure we used is quite satisfactory. It can
NULL * 0.6798 29 TRUE recognize both the known and unknown images in the
database in various conditions with accuracy 100%. We think
1.0e+005 when we use a huge database such as thousand number of
NULL * 0.3922 21 TRUE images or more, then some misdetection (2%-5%) may
1.0e+005 happened in the recognition procedure. Our future work may
NULL * 0.1790 23 TRUE include an improvement to the eigenfaces technology and
recognition procedure for live video stream.
1.0e+005
NULL * 0.2576 1 TRUE

48
 International Journal of Computer Applications (0975 – 8887)
Volume 50 – No.10, July 2012

7. REFERENCES [10] Harmon, L. D., and Hunt, W. F., "Automatic recognition

[1] F.Galton, “Personal identification and description 1,1 of human face profiles", Computer Graphics and Image
Nature, pp.173-177,21 June1988 Processing, Vol. 6, pp. 135-156, (1977).

[2] Sir Francis Galton, Personal identification and [11] Harmon, L. D., Khan, M. K., Lasch, R., and Raming, P.
description-II”, Nature 201-203, 28 June 1988 F., “Machine identification of human faces”, Pattern
Recognition, Vol. 13(2), pp. 97-110, (1981).
[3] Goldstein, A. J., Harmon, L. D., and Lesk, A. B.,
Identification of human faces", Proc. IEEE 59, pp. 748- [12] Kaufman, G. J., and Breeding, K. J, “The automatic
760, (1971). recognition of human faces from profile silhouettes”,
IEEE Trans. Syst. Man Cybern ., Vol. 6, pp. 113-120,
[4] Haig, N. K., "How faces differ - a new comparative (1976).
technique", Perception 14, pp. 601-615, (1985).
[13] Wu, C. J., and Huang, J. S., "Human face profile
[5] Rhodes, G., "Looking at faces: First-order and second recognition by computer", Pattern Recognition, Vol.
order features as determinants of facial appearance", 23(3/4), pp. 255-259, (1990).
Perception 17, pp. 43-63, (1988).
[14] Kerin, M. A., and Stonham, T. J., "Face recognition
[6] Kirby, M., and Sirovich, L., "Application of the using a digital neural network with self-organizing
Karhunen-Loeve procedure for the characterization of capabilities", Proc. 10th Int. Conf. On Pattern
human faces", IEEE PAMI, Vol. 12, pp. 103-108, Recognition, pp.738-741, (1990).
(1990).
[15] Nakamura, O., Mathur, S., and Minami, T.,
[7] Sirovich, L., and Kirby, M., "Low-dimensional "Identification of human faces based on isodensity
procedure for the characterization of human faces", J. maps", Pattern Recognition, Vol. 24(3), pp. 263-272,
Opt. Soc. Am. A, 4, 3, pp. 519-524, (1987). (1991).
[8] Terzopoulos, D., and Waters, K., "Analysis of facial [16] Turk, M., and Pentland, A., "Eigenfaces for recognition",
images using physical and anatomical models", Proc. 3rd Journal of Cognitive Neuroscience, Vol. 3, pp. 71-86,
Int. Conf. on Computer Vision, pp. 727-732, (1990). (1991).
[9] Manjunath, B. S., Chellappa, R., and Malsburg, C., "A [17] Yuille, A. L., Cohen, D. S., and Hallinan, P. W., "Feature
feature based approach to face recognition", Trans. of extraction from faces using deformable templates", Proc.
IEEE, pp. 373-378, (1992). of CVPR, (1989).

49

View publication stats