International Journal of Control, Automation, Communication and Systems (IJCACS), Vol.1, No.

1, January 2016

PERFORMANCE ANALYSIS BETWEEN PCA

AND ICA IN HUMAN FACE DETECTION
Md. Shahjahan Kabir1, Md. Solaiman Hossain2 and Dr. Rabiul Islam3
1, 2
Department of CSE, Dhaka Int. University (DIU), Dhaka, Bangladesh.
3
Associate Professor, Dept. of CSE, RUET, Rajshahi, Bangladesh.

ABSTRACT
In the context of Face Recognition the paper compares between Principal Component Analyses (PCA) and
Independent Component Analysis (ICA). In the psychological and algorithmic literature Principal
Component Analyses (PCA) and Independent Component Analyses (ICA) are the basis of numerous studies.
Classical technique in statistical data analyses is called Principal Component Analyses (PCA) and
technique of array processing and data analysis is called Independent Component Analysis (ICA). This
paper gives the better concept how each algorithms are worked. This concept helps to advance level
experiment. Besides, those method is so strong others system and algorithm. In both case, ICA performs
good but not as good as PCA. For Face Recognition Principal Component Analyses basis algorithm
represents intelligent suction of a random search within a short time and its can detect to solve the
problem.

KEYWORDS
Face Recognition, ICA, PCA, JAFFE database.

1. INTRODUCTION

Face Recognition is important part of computer vision. Besides the last some years Face
Recognition is a powerful application. There are many commercial systems for human face
detection is available now. For advance research efforts have been focused on video based face
modelling or tracking, recognition and system integration. It’s an important part of ability of
human appreciation system. During the past years, Face Recognition has received increased
evolution. Compression data is mandatory to use, Face Recognition better for this purpose. It has
special technically views. Images can be compressed as structural features such as contours and
regions. [1] It can be access through the human computer interface (HCI) in crowd surveillance.
The earliest system on Face Recognition can be traced back at least to the 1950s in psychology
and to the 1960s in the in the engineering literature. As perfect by the algorithms, this paper
compares Principal Component Analysis (PCA) to Independent Component Analysis (ICA).
Advantages of using Principal Component Analysis are data can be compressed image without
harm any data. PCA can outperform over many other techniques when the size of database is
small. In proposed algorithm the database is sub grouped using some features of interest in face.
[2] The performances of appearance based statistical methods such as PCA and ICA are tested
and compared for the recognition of coloured faces images in Face Recognition system. [3] The
performance of PCA depends on the task statement, the subspace distance metric. The
performance of ICA depends on the task, so the algorithm is used to approximate ICA. The goal
of this paper is to compare a canonical subspace projection technique Principal Component
DOI: 10.5121/ijcacs.2016.1104 25
International Journal of Control, Automation, Communication and Systems (IJCACS), Vol.1, No.1, January 2016

Analysis (PCA) to a newer technique, Independent Component Analysis ICA. [4, 5, and 6]
Especially in Face Recognition PCA is popular method and ICA is originally developed for
separating mixed audio signals into independent sources. In the Figure 1, the learning stages for
Face Recognition System are shown. Initially, it collects face images as an input from the
database. After that, feature extraction is happened by the processing of face images. In the next
step, learning algorithm is used and finally kept images in training database for the testing stage.
These steps are shown below.

Face collection & input

Pre-processing of face image
Feature Extraction
Learning Algorithm
Saving the image in training database for the testing stage.

Figure 2. Learning stage for Face Recognition.

In the Figure 2, at first take input of face image for matching. Then processes the face image for
feature extraction. Finally applied the recognition algorithm and provide the output result for
successfully recognised or unsuccessfully recognized. These steps are shown below.

Face (input)
Pre-processing of the face image
Feature Extraction
Recognition Algorithm
Success or Failure (output)

Figure 2. Testing stage for Face Recognition.

But actually the percentage of Face Recognition system of ICA is still not so good. For this
reason, Face Recognition system is still now researched. In this paper, the algorithm of PCA and
ICA are proposed. PCA is more powerful and sharp than ICA. Due to the percentage reason, the
proposed algorithm is increasing noise free system for Face Recognition, which is more standard
than previous.

26
International Journal of Control, Automation, Communication and Systems (IJCACS), Vol.1, No.1, January 2016

2. PCA & ICA

For PCA all images must be same size and resolution and one image is treated as one vector.
Gallery images represents as matrix and images are set are stored in a single matrix. The average
image has to be calculated and then subtracted from each original image. Covariance matrix is
calculated by eigenvectors and Eigen values. These eigenvectors are called Eigen face and its
remove the unsuccessful information. Besides, it fester the face structure into the uncollected
component. A Block Diagram of Face Recognition System using Principle component analyses
are shown in Figure 3:

Figure 3. Block diagram of PCA

ICA of a random vector searches for a linear transformation which minimizes the statistical
dependence between its components. ICA basis vectors are more spatially localized than their
PCA counterparts Evaluations on localized recognition tasks, such as recognizing expressions,
may produce significantly different results. [7] Basically ICA is also good for Face Recognition
system. Local representation provides robustness to partial occlusions and local distortions
because successful Face Recognition can be achieved by representing some important facial parts
that correspond to feature regions such as eyes, eye brows, nose and lips. This “recognition by
parts” paradigm has been popular in the object recognition research because the approach can be
successfully applied to the problem of object recognition with occlusion. [8] The first ICA
method computes Independent Components by maximizing non-Gaussianity of whitened data
distribution using a kurtosis maximization process. The kurtosis measures the non-Gaussianity
and the sparseness of the face representations. [9] Here this system, need to an input image. After
the input image it’s calculate the Eigen values of training images. Then select a test image or
original image. Apply ICA algorithm in the test image. Then get a mean image from the test
image. If it matches the training images, the result found otherwise not shown in Figure 4.

27
International Journal of Control, Automation, Communication and Systems (IJCACS), Vol.1, No.1, January 2016

Figure 4. Block diagram of ICA

3. PROXIMITY OF PCA

For Face Recognition PCA is probably the most widely subspace technique. For PCA,
Obtain M training images I1 , I 2 ,.......I M because the images are cantered.

On this formula below, each image Ii represented as vector

a 11 a 12 a1N  a 11 
 a 1 N 
a 21 a 22 a 2 N

 a 2N  T
i
 
a N 1 a N 2 a NN N* N  a NN 

Now find out the average face vector.

∑
M
Ψ = 1/ M i =1
Ti (1)

28
International Journal of Control, Automation, Communication and Systems (IJCACS), Vol.1, No.1, January 2016

Figure 5. Mean Image

Subtract the mean face from each face vector, to get a set of vectors.

(2)
Φ i = Ti − Ψ
Find the covariance matrix C=AAT Where,

A =[Φ1, Φ2,......ΦN ] (3)

2 2 2
But C=N * N and A=N *M , here C is huge.
Then find out the, L=ATA=M*M (4)
Now can get V eigenvector and then calculate the eigenvector of C using this equation (5),
Ui = AVi (5)

Figure 6. Eigen Faces

Weight calculation for each training images using this formula (6),

W j = U Tj Φ i (6)

At last calculate the individual weight vector for each training images using the equation (7),

 w1  (7)

Ω i =  w2 
 wN 

For test image also calculated, Ω. At the last step distance measured from test image to every
training image using Euclidean or City blocks distance. The result is found whose distance is
minimum.

Eudist = (test − train ) 2 (8)

29
International Journal of Control, Automation, Communication and Systems (IJCACS), Vol.1, No.1, January 2016

Figure 7. Face Recognized using PCA

4. PROXIMITY OF ICA

Independent Component Analysis (ICA), a generalization of PCA, is one such method. At first it
has number of source K. From the K source, here found the training image which number is n.
Here, X=AS. The De-mixing matrix is W.W is an inverse of Recovered the ICA, U=WX.

W is comes from the Informix algorithm proposed. W is selected randomly. W is increased by

this algorithm is given bellow.

fi (U)=1/1+e-U (9)

Where,
Yi = f ''i (U i ) / f 'i (U i ) (10)

∆W = ( I + Y 'U T )W (11)

Here the identity matrix and the logistic transfer function gives,

Yi = (1 − 2Yi ) (12)

X are the training set of images .The whitening matrix is comes from the equation (13),
(13)
W z = 2 * (cov( X )) ( − 1 / 2 )

When the inputs to ICA are the “sphere” data, the full transform matrix Wi is the product of
the sphering matrix and the matrix learned by ICA,

(14)
WI = W *Wz

The PC representation of the set of zero-mean images in X based on Pm is defined as,

Rm =XPm (15)

30
International Journal of Control, Automation, Communication and Systems (IJCACS), Vol.1, No.1, January 2016

For train image,

B = RmWI−1 (16)

For test image,

Rtest = XtestPm Btest = RtestWI−1

Figure 8. Test image for ICA

Calculate the cosine matrix between test and training set by using this equation (17),
c = Btest * Btrain / Btest Btrain (17)

Figure 9. Cosine matrix of an image

Euclidean Distance measure between training and testing image using the cosine angle c=α. same
as PCA.

d = ( test − train ) 2 − 2 cos( c ) (18)

Figure 10. Face Recognised by ICA

31
International Journal of Control, Automation, Communication and Systems (IJCACS), Vol.1, No.1, January 2016

5. PERFORMANCE & RESULT ANALYSIS

For the data Analysis JEFFE database are used. For the analysis, machine has got random results
for random number of images. For ICA and PCA, the results are shown below table 1 and table 2.

Table 1: Competency Calculation for ICA

No. of Face 15 25 35 45
Image
Successfully Recognized 14 23 31 40
Face Image

Unrecognized Face Image 1 2 4 5

COMPETENCY (%) 94% 92% 88% 89%

Table 1 showed that 90.75% Competency of Face Recognition System by using ICA
(Independent Component Analysis).

Table 2: Competency Calculation for PCA

No. of Face 15 25 35 45
Image

Successfully Recognized 15 23 33 43
Face Image

Unrecognized Face Image 0 2 2 4

COMPETENCY (%) 100% 92% 94% 93%

Table 2 showed that 94.75% Competency of Face Recognition System by using PCA (Principal
Component Analysis). If Plot those data into a graph then the graph is shown below Figure 11
and also an efficiency calculation graph are shown Figure 12.

32
International Journal of Control, Automation, Communication and Systems (IJCACS), Vol.1, No.1, January 2016

3. CONCLUSIONS

Face Recognition System using the concept of Principle Component Analysis (ICA) and
Independent Component Analysis has been discussed. The maximum ability for Face Recognition
System by using Principal Component Analysis (PCA) is 94.75% and the minimum efficiency for
Face Recognition System by using Independent Component Analysis is 90.75%. For this
algorithm and database system another important part is camera, cause if good resolution digital
camera or good resolution analogue camera is used, the results could be increased and it will give
better result. Otherwise it can’t carry good result for this system. Principal Component Analysis
(PCA) is very safe and better system for search an image from training database. On the other
hand Independent Component Analysis (ICA) is good system but critical for comfortable uses.

33
International Journal of Control, Automation, Communication and Systems (IJCACS), Vol.1, No.1, January 2016

REFERENCES

[1] Mahima Agrawal1,Shubangi.D.Giripunje2,P.R.Bajaj3, “Recognizing Facial Expression using PCA

and GeneticAlgorithm”, International Journal of Computer & Communication Technology, April
2011.
[2] Bruce A. Draper,a,* KyungimBaek,b Marian Stewart Bartlett,c and J. Ross Beveridge (Computer
Vision and Image Understanding 91 (2003) 115–137)
[3] Önsen TOYGAR Adnan ACAN ,ǁFace recognition using PCA,LDA and ICA approaches on colored
imagesǁ, Journal Of Electrical and Electronics Engineering, vol-13,2003
[4] Turk, M. and A. Pentland, Eigenfaces for Recognition. Journal of Cognitive Neuroscience, 1991.
3(1): p. 71-86
[5] Comon, P., Independent component analysis; A new concept? Signal Processing, 1994. 36(3): p. 287-
314.
[6] Bell, A.J. and T.J. Sejnowski, An informationmaximization Approach to Blind Separation and Blind
Deconvolution. Neural Computation, 1995. 7(6): p. 1129-1159.
[7] Donato, G., et al., Classifying Facial Actions. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 1999. 21(10): p. 974-989
[8] A. P. Pentland, “Recognition by parts,” IEEE Proceedings of the First International Conference on
Computer Vision, pp. 612-620, 1987.
[9] A. J. Bell and T. J. Sejnowski, “The Independent Components of Natural Scenes are Edge FILTERS,”
VISION RESEARCH, VOL. 37, NO. 23, PP. 3327-3338, 1997.

Authors

Md. Shahjahan Kabir is a faculty member of CSE in Dhaka International University. He has got B.Sc.
from RUET. He has online publications in the different international journals.

Md. Solaiman Hossain is a student of B.Sc. at CSE in Dhaka International University.

Dr. Rabiul Islam is currently an Associate professor at CSE in RUET. He has already completed his
B.Sc., M.Sc. and PhD. from RUET. He has lots of research and publications. He has already published 65
online publications in different international journals.

34