3rd International Conference on Advanced Computing & Communication Technologies, November 08-09, 2008, APIIT, Panipat, India.
A Novel Approach for Recognition of Human Faces Jay Shankar Prasad*, G C Nandi**
Indian Institute of Information Technology, Allahabad, India. Email: *
[email protected] , **
[email protected] lighting condition, variations in pose, variation in facial expression the recognition gets affected [3, 4, and 7]. In this paper we compare the effect of no of eigen vectors , resolution, No. of training samples, and role of classifier using PCA in recognition of human faces .Face data set is originally in color but converted to gray scale and black and white to perform several experiments. Yongsheng [1] shown that variations have greater impact on the facial recognition. They used Line map edge approach for face image recognition which extracts lines from face edge map as geometrical feature. The recognition rate achieved by them is fairly good. They found that time complexity of the LEM algorithm is more as compared to Eigen face method but the space complexity of the eigen face method is less than LEM method. Purnell and Botha [7] experimented PCA on heterogeneous population they performed experiment on the database which is of different subcontinent of the world, they found that almost 65 to 75 percent eigen vectors are sufficient to classify the human faces. The method applied is segmentation but the drawback observed is that if either of the eye is not present in the image it is not able to recognize the correct faces. Classifier used is nearest neighborhood and neural network. Neural network classifier is time consuming in training. From both the classifier they found that while increasing the number of training samples recognition performance were improved .When the number of training samples is restricted nearest neighborhood performs well. The neural network performance is good when training samples are large and numbers of eigen vectors are also large. The recognition achieved is below to 83 percent. Komleh et.al. [2] Gave expression invariant solution for the face recognition based on fractal features. The algorithm performs very well for small database but the performance degrades while increasing training samples within database. Zhao and Yuan [3] also used a variant of PCA to improve the recognition. They found that singular value decomposition of features can be used with PCA for face recognition. Boom et.al [4] investigates that how image resolution affects the result of a face recognition system. They found that PCA approach can works even in low resolution image of 32x32 pixels. They also studied the effect of resolution and PCA components .Increasing the number of components does not improve the recognition if the resolution is low. In a low quality face
ABSTRACT: This paper attempts to describe the effect of various parameters on the facial image recognition. We considered method which is concerned with representation and recognition of images using principal component analysis. The parameter which plays the role in the representation is resolution of image and pose of human face. The recognition process gets affected with eigen vector, no of training and testing sample, type of classifier etc. We use Euclidean distance and Mahalanobis distance classifier for our experiments. We found very promising result using Euclidean classifier .The result of our system are optimal for 64x64 pixel image. Key terms: Face recognition; PCA; Euclidean distance classifier; Mahalanobis distance classifier
1.
INTRODUCTION
Face recognition is used to identify a particular person from a video sequence or still images through some technique. The images are stored in the database .Computerized human face recognition is an interesting area of research since last 25 years [8]. It has several practical applications in security and access control, surveillance, interpreting human actions, behaviors. The precise identification system is in demand. Password identification and token identification system is available for security and access control. Because these systems are vulnerable to forgery and theft .Apart from it also due to lapses in user’s memories, biometric identification systems is used .These biometric systems are based on pattern recognition techniques to identify human faces by their characteristics are attracting significant interest. Twenty five years ago, the problem of face recognition was considered among the hardest in artificial intelligence and computer vision. This paper presents novel approach for recognition of human faces. Previous attempt to recognize human faces are based on geometrical and template based methods. Here we are going to recognize human faces based on Principal component analysis [7, 8, and 9]. The problem with face recognition technique is that due to variation in
258
3rd International Conference on Advanced Computing & Communication Technologies, November 08-09, 2008, APIIT, Panipat, India.
data set face is not correctly detected and recognition is low. Nefian and Hayes III [6] used Hidden markov model for recognition of human faces. They extracted two dimensional discrete cosine transform for feature vectors. The experiment was performed while varying the facial expressions, hair style, Head orientations and eye wear (glass /no glass). Nazeer, Omar and Khalid [5] used artificial Neural Network approach for recognition of human faces through photometric normalization technique. Euclidean and normalization correlation classifier were used. For extracting the feature PCA and LDA methods were used by them. The result of PCA was fairly good than LDA. The neural network classifier gives the minimum recognition rate among Euclidean distance, normalization correlation and neural network classifier. 2.
These difference images are used for finding out the covariance matrix for our dataset. The covariance signifies how much the data sets correlates. C=
=
=AAT
n (8) From the covariance matrix C the eigen value and eigen vector is obtained using equation (1). These eigen vectors determine linear combination of the M training set face image to form the eigen faces Ul where , l=1,.., M. (9) Ul= Thus eigen value i represent the variance of the corresponding face image set along the axis described by i. By selecting the eigen vector with largest eigen value as our basis, we are selecting the dimensions which can express the greatest variance in facial image or in other words the dominant modes of face space.
METHODOLOGY
We use the following framework for training and recognition of human faces: Image Acquisition Face Image Pre- Processing Normalised Face Image Feature Extraction
Classification
Training /Testing
Fig 1: Training Set of facial images
Face Database
Each image is transformed into a vector of size N and placed into the set. From training set (e.g. Fig 1) mean image has been obtained using equation (6) the image obtained is presented in fig (2). The difference between input and mean image is calculated. The covariance matrix of the image is found with the help of equation (8). From the covariance matrix the eigen vectors and eigen values are computed. The eigen vectors are arranged according to descending order of eigen values. Normalization of image has been done to overcome the uncontrolled variations in lighting conditions. If A is a matrix of n x n size then X is eigen vector and λ is the corresponding eigen value then AX=λX (1) Equation (1) can be true only if det (2) A matrix is symmetric if or = (3) It is said to be orthogonal if =I (4) The matrix is normal if A (5) Let the training set of facial image be , , the average set can be defined as Y=
Fig 2: Mean face obtained from Fig (1) Training images
Fig 3: All Training set images are converted into Eigen Faces
(6)
An image of 128 x 128 pixel size can be represented as a point in a 16384 dimensional space. Using this method a face can be accurately represented with very few coordinates as 8. It means that the earlier face was taking
Each face differs from the average face by the vector -Y (7)
259
3rd International Conference on Advanced Computing & Communication Technologies, November 08-09, 2008, APIIT, Panipat, India. Fig 4: Euclidean distance for a test image
16384 byte of memory space to represent it now it takes only 8 bytes. It makes the work of face recognition simple. In our experiment we created 256x 256, 128 x 128, 64 x 64, 48 x 48, 32 x32, 16 x 16 and 8 x 8 resolution face image. From L eigen vectors Vl only L’ eigen vectors are selected and known as principal component which have the highest eigen value association. Training of the face data is finished after obtaining the eigen vectors of all eigen faces. Fig 3 shows the Eigen faces obtained from the training images. 3.
4.
In this section we carry out some experiments to study the effects of different parameters on the recognition. For all the experiments the methodology is applied as mentioned in the previous section. Database contains the images of 40 male and 40 female face images. The images are taken under various lighting conditions with front and side poses. All the experiments is performed for color, gray scale and black and white images.
CLASSIFICATION OF FACES
To classify a new face Fnew to one of the classes of known faces, new face image is transformed into its eigen face components. For some test image Fnew , its projection on axis I- Wk is given by Wk= UkT(F new – Y), k= 1... no of eigen faces used for face recognition .Vector Fnew T is the weight vector of the image I which contains the projections of Fnew onto each of the dominant eigen vector and can be expressed through following equation ZnewT= [ w1 w2 … wm’]. The Euclidean distance between two weight vectors d(Zi, Zj) provides a measure of similarity between the corresponding image i and j [7, 8]. In the same way Mahalanobis distance measure [9] also gives the similarity between image i and j. If the distance exceeds some threshold value T set by us we say Fnew is no face at all. The Mahalanobis distance from a and group of values with mean covariance matrix Σ for a vector x=(x1, x2, x3,…, xp)T is defined as DM(x) =
EXPERIMENTS, RESULTS AND ANALYSIS
Experiment 1: We made a face training data set which contains frontal faces of 40 individuals. The test set contains the frontal faces but different image than training images. We changed the no of training and testing image and analyzed the result. The Resolution of the image we take is 256X256 pixel size. We took 70 % of eigen vector as principal component. The experiment was performed first on Color image, then on gray scale image and black & white image. The classifier used by us is Euclidean Distance. As we boost the number of training samples the recognition improves. But we cannot contain too many samples in our data base because it increases the training time. Fig (5) shows that recognition is linearly dependent on No. of sample taken for training. Fig (6, 7) shows how no of testing sample also affect the recognition rate. In fig (6) we had taken only 30 % test sample, all the constraints are same. Fig (7) depicts the recognition rate when we had taken 60 % testing samples. The recognition rate falls (fig 7) as compared to previous (fig 6).
(12).
Euclidean distance between points X and Y in n space is defined as (13) 100
The Euclidean distance calculated using equation (13) for a test image is like in fig (4) and we are selecting the minimum distance for matching a particular image with test image. 9200
60
40 100 20
80
0
Euclidean Distance
8800
90
85
80
20 75
70 Testi ng Im age (%)
8600
8400
65
60
0
Trg image vs Testing Img vs Color Trg image vs Testing image vs Gray Trg image vs Testing image vs B & W
8200
8000
7800
40
Tr ai ni ng
95
Im ag e
60
9000
(% )
) Recognition (%
80
Fig 5: Effect of Training Samples and Testing Samples on face recognition 0
5
10
15
20 25 Training Image
30
35
40
260
3rd International Conference on Advanced Computing & Communication Technologies, November 08-09, 2008, APIIT, Panipat, India.
100
70 80
50
Recognition (%)
Recognition(%)
60 60
40
20
0 20
40
60
40
30
20
80
Training Image (%)
10
Color Gray B&W
0 10
Fig 6: Effect of Training samples on the recognition of faces while 30 % test samples taken.
20
40
50
60
70
80
Eigen Vector (%) Color Gray B&W
Fig 9: Recognition using Mahalanobis distance classifier
80
60
R ecognition(% )
30
Experiment 3:
40
Here we considered various resolution images to study the effect of resolution of images on the recognition. We use 75 % training samples and 25 % testing samples for this experiment. We use 70 % eigen vector for this experiment. Fig 10 and 11 shows the effect of resolution on the recognition. The best result obtains from 64x64 pixel image resolution for both the classifier. The number of pixels displayed per unit of printed length in an image, usually measured in pixels per inch (ppi).The amount of detail in an image depends on its pixel dimensions, while the image resolution controls how much space the pixels are printed over.
20
0 20
40
60
80
Training Image (%) Color Gray B&W
Fig 7: Effect of Training samples on the recognition of faces while 60 % test samples taken.
Experiment 2: We changed the no of eigen vectors while we fixed the training samples at 75 % and testing samples at 25 %, all images are 256 x256 pixel resolutions. The classifier used by us is Euclidean distance and Mahalnobis distance measure. Figure (8) shows the recognition of faces when we are using the Euclidean distance classifier. Figure (9) shows the recognition using Mahalanobis distance classifier. From figure 8 and 9 it is clear that Euclidean distance classifier is advantageous over Mahalanobis distance.
84 82
Recognition (%)
80 78 76 74 72
100
70
R e c o g n itio n(% )
80
68
60
66 256 x 256 128 x 128 64 x 64
48 x 48
32 x 32
16 x 16
8x8
40
Resolution in pixels 20
0 10
20
30
40
50
60
70
80
Eigen Vector(%)
Color Gray B&W
Color Gray B&W
Fig 10 : Effect of resolution on the recognition using Euclidean distance classifier
Fig 8: Recognition using Euclidean Distance Classifier
261
3rd International Conference on Advanced Computing & Communication Technologies, November 08-09, 2008, APIIT, Panipat, India. [7] 70
[8] [9]
Recognition (%)
65
60
55
50
45
40 256 x 256 128 x 128 64 x 64
48 x 48
32 x 32
16 x 16
8x8
Resolution in pixels Color Gray B&W
Fig 11: Effect of resolution on the recognition using Mahalanobis distance classifier
5.
CONCLUSION
This paper describes an PCA based approach for face recognition that uses efficient set of eigen vectors based on the extraction of eigen values and eigen vectors. The experimental result shows Euclidean distance classifier is superior to Mahalanobis distance classifier. Resolution also plays a vital role in the recognition process. The inclusion of large number of high order eigen vectors increases the computational complexity of the system, the performance of the system does not improve simply through increasing the number of eigen vectors. To improve the recognition rate some other type of features may be taken and the hybrid classifier can also be used. 6. [1]
[2]
REFERENCES
Yongsheng Gao, “Face Recognition Using Line Edge Map”, IEEE Transaction on Pattern Analysis and Machine Intelligence, Vol 24 No. 6, pp .764-779, June 2002. Hossein Ebrahimpour-Komleh, Vinod Chandran, Sridha Sridharan, “Robustness to Expression Variation in Fractal –based Face recognition,” proc. IEEE conf. on Signal Processing and its Applications (ISSPA), pp 359-362, August 2001.
[3]
[4] [5]
[6]
Haitao Zhao, Pong Chi Yuen, “A Novel Incremental Principal Component Analysis and Its Application for Face Recognition”, IEEE Trans. on Man Machine and Cybernetics, pp. 873-885, August 2006. B.J. Boom, G.M. Beumer, L.J. Spreeuwers, R. N. J. Veldhuis, “The Effect of I mage Resolution on the Performance of a Face Recognition System”, IEEE conf. ICARCV 2006 Shahrin Azuan Nazeer, Nazaruddin Omar and Marzuki Khalid, “Face Recognition System using Artificial Neural Networks Approach”, IEEE - ICSCN 2007, MIT Campus, Anna University, Chennai, India”,pp.420-425,Feb. 22-24, 2007 Ara V Nefian, Monson H Hayes III, “Hidden Markov models for face recognition”, Proc. International Conf. on Acoustics, Speech and Signal Processing, pp. 2721-2724, 1998
262
D.W. Purnell, C. Nieuwoudt and E. C. Botha, “Face recognition in a heterogeneous population”, Pattern Recognition Letters, pp. 594-599, Sep 1998. M. A. Turk and A. P. Pentland, “Face recognition using eigenfaces,” Proc. Int. Conf. on Pattern Recognition, pp. 586-591, 1991. Ian Craw, Nicholas Costen, Takashi Kato, and Shigeru Akamatsu, “How Should We Represent Faces for Automatic Recognition?”, IEEE Trans. on Pattern Analysis and Machine Inetlligence, Vol. 21, NO. 8, pp. 725-736, August 1999