Face Recognition Using Trichotomic Combination Of SVD, DF-LDA And LPP

Shylaja S S1, K N Balasubramanya Murthy, S Natarajan Pritha D N, L Savitha Department of Information Science and Engineering, P E S Institute of Technology, INDIA

Abstract: One of the challenges the face recognition application is facing today is that of the high dimensionality of multivariate data. In this context, this paper proposes to compare the performance of a triumvirate combination of linear dimensionality reduction techniques namely Singular Value Decomposition (SVD) which maximizes the variance of the training vectors, Direct Fractional Linear Discriminant Analysis (DFLDA) that maximizes the "between-class" scatter while minimizing the "within-class" scatter and Locality Preserving Projection (LPP) which preserves the local features those unique from its nearest neighbors. The amalgamation containing different ratios is chosen from the features extracted by the three independent techniques mentioned above. Original Face space is projected onto the manifold of chosen basis. The weights obtained from these projections for the probe set are compared with that of the query image using the mean distance classifier. The proposed method has been tested on YALE dataset and the combination in the ratio 3:2:5 showed significant improvement in the efficiency of recognition, with a calculated accuracy of 92.7% on a test set of 165 images Keywords: Pattern Recognition, Classification, Dimensionality Reduction, SVD, DFLDA, LPP, Receiver Operating Characteristics (ROC).

faces from LDA technique[4,5] extracts discriminating features between classes and is found to perform better for large data sets. Its shortcoming is that of Small Sample Space (SSS) problem. LPPs[3] are linear projective maps that arise by solving variational problem that optimally preserves the neighborhood structure of the data set. In this paper the best combination of features obtained from Eigen, Fisher and Laplacian faces is estimated to improve the efficiency of face recognition task. Two trails are performed for 1/3rd and 2/3rd reduction in dimension. For every trial, several iterations are performed with different ratios which will eventually converge to the right basis. This basis describes holistic, discriminant as well as unique features for the face space. Projection of the face space onto the subspace spanned by this basis is used in classification. Rest of the paper is organized as follows: section 2 covers the dimensionality reduction; sections 3, 4, 5 describe SVD,DF-LDA & LPP; proposed methodology is discussed in 6,;experimental results are illustrated in section 7 and section 8 focuses on conclusions & suggestions for further work. II. RELATED WORK

2.1 DIMENSIONALITY REDUCTION

In statistics, dimension reduction is the process of reducing the number of random variables under consideration, RN RM (M<N) and can be divided into feature selection and feature extraction. Feature selection is choosing a subset of all the features and feature extraction is creating new features from existing ones. In either case, the goal is to find a low-dimensional representation of the data while still describing the data with sufficient accuracy.

I.

INTRODUCTION

Face recognition is a pattern recognition task performed specifically on faces. It can be described as classifying a face as either "known" or "unknown". Computational models of face recognition must address several difficult problems. The problem of dimensionality reduction arises in face recognition because a m X n face image is reconstructed to form a column vector of mn components, for computational purposes. As the number of images in the data set increases, the complexity of representing data sets increases. Analysis with a large number of variables generally consumes a large amount of memory and computation power. To overcome the curse of dimensionality, various dimensionality reduction techniques were formulated. The linear methods include PCA, LPP, LDA, SPP[1,2,3] and non linear methods[6,7] include ISOMAP & Eigenmaps. Techniques such as PCA have the advantage of capturing holistic features but ignore the localized features. Fisher
1

2.2 SINGULAR VALUE DECOMPOSITION (SVD)

Singular value decomposition (SVD) is an important factorization of a rectangular real or complex matrix, with many applications in signal processing and statistics. As applied to face recognition this technique is used to extract the holistic global features of the training set. The mathematical model followed is described in the following steps. Let A is m X n real matrix and N=ATA

Corresponding Author: shylaja.sharath@pes.edu

978-1-4244-5586-7/10/$26.00
C

2010 IEEE

246
Volume 1

R denotes the range space and N denotes the null space of a matrix. orthonormal basis v i 1 i are sought for RAT where is the rank of RAT& ui 1 i for RA such that , Avj=sju and ATuj=sjvj 1j (2) (1)

2.4 LOCALITY PRESERVING PROJECTIONS

LPP[3] shares many of the data representation properties of nonlinear techniques such as Laplacian Eigenmaps or Locally Linear Embedding. Yet LPP is linear and more crucially is defined everywhere in ambient space rather than just on the training data points. It builds a graph incorporating neighborhood information of the data set. Using the notion of the Laplacian of the graph, transformation matrix is computed which maps the data points to a subspace. This linear transformation optimally preserves local neighborhood information in a certain sense. The representation map generated by the algorithm may be viewed as a linear discrete approximation to a continuous map that naturally arises from the geometry of the manifold. The mathematical model describing LPP procedure is given the following paragraph. A weight matrix using k nearest neighbors is constructed such that Wij =1 if xi is among k nearest neighbors of xj. Eigenmaps are obtained from the generalized eigenvector problem as ALAT a = ADAT a where D is a diagonal matrix whose entries are column or row, since W is symmetric sums of W, Dii = jWji., L = D -W is the Laplacian matrix is equivalent nonlinear Laplace Beltrami opearator. The ith column of matrix A is xi. Let the column vectors a0; _ _ _ ; al-1 be the solutions of equation (), ordered according to their eigenvalues, in ascending order Thus, the embedding is as follows: yi = ET xi; E = (a0; a1; _ _ _ ; al-1) where yi is a l-dimensional vector, and E is a n x l matrix.. The yi represent the Laplacian faces. III. PROPOSED METHOD The method proposed uses different ratios of features obtained from SVD, DF-LDA & LPP Techniques. The first step will normalize images of the training set to compensate for the illumination effects. These processed images undergo dimensionality reduction using each of the methods mentioned above. The three independent sets of dimension reduced vectors so obtained are used in different proportions to arrive at a suitable basis which will enhance the efficiency of recognition. Two trials have been used with one for 1/3rd dimensionality reduction and another for 2/3rd reduction. In each of the trials, several iterations are performed by taking different combinations of the feature vectors. The iterations will converge when the desired precision of recognition rate is obtained. 3.1 Preprocessing 3.1.1 The Face Space

Advantages of having such a basis are that geometry becomes easy and gives a decomposition of A into oneranked matrices. Combining the equations (1) & (2) A=sjvjujT 1 j The resulting ui span the Eigen subspace.

2.3 DIRECT FRACTIONAL STEP LINEAR DISCRIMINANT ANALYSIS

The DF-LDA[4] is a linear pattern recognition method. Compared with nonlinear models, a linear model is rather robust against noises and most likely will not over fit. Although it has been shown that distribution of face patterns is highly non convex and complex in most cases, linear methods are still able to provide cost effective solutions to the FR tasks through integration with other strategies, such as the principle of divide and conquer, in which a large and nonlinear problem is divided into a few smaller and local linear sub problems. Let and denote the between- and withinclass scatter matrices of the training image set, respectively. LDA-like approaches such as the Fisherface method find a set of basis vectors, denoted by that maximizes the ratio between and

3 The maximization process in (3) is not directly linked to the classification error which is the criterion of performance used to measure the success of the Face Recognition procedure. Thus, the weighted between-class scatter matrix can be expressed as: (4) where , is the mean of class , is the number of elements in , and

is the Euclidean distance between the means of class i and j .

247
Volume 1

For the recognition task , each m X n Ii image in the training set is transformed into a column vector of mn components.a matrix S (mn X M) is constructed such that S =[ I1 I2 . . . IM] where M is number of face images in training set It is found that all N vectors are linearly independent, which implies that the range space of matrix S is the entire region spanned by the columns of S. i.e Range space of S R(S)=[ S] 3.1.2 Normalization Images are normalized to reduce illumination effects and lighting conditions as, Ai= i=1, 2,, and

i.e. the orthogonal projection of a face vector on each basis vector. Recognition Task: Unknown probe face is normalized and projected onto the subspace to get weight for the probe image. . Mean distance classifier . is used And if , where is a threshold chosen heuristically, then the probe image is recognized as the image with which it gives the lowest score. If however then the probe does not belong to the database. Deciding on the Threshold: In choosing the threshold , a set of 150 known images other than the ones in dataset are taken, then minimum distance is calculated and stored in vector . The threshold is chosen as, where,

3.2 Basis Selection 3.2.1 Basis from SVD If A is the face Space, then x vectors are obtained from [X1..Xx] = <1..x>(A-1UD), where U & D are the unitary and diagonal matrices SVD of A. 3.2.2 Basis from DF-LDA The set Y vectors are chosen by the given equation [y1..yy]= <1..y>(UTSTOTUT), where STOT is the sum of between and within class scatter matrices, U is a diagonal matrix from eigen values and vectors. 3.2.3 Basis from LPP Locality information can be preserved by the following transformation on A, the input face space [z1 . Zz]= <1..z(AT L A), where L =D-W gives the Laplacian matrix. D is the diagonal matrix and W is the weight matrix of the K nearest neighbors clustering. Basis for the face space is obtained as, Such that, and M is dimension of original face space. Projection onto reduced subspace: Each face in the training set can be represented as a linear combination of these vectors, ui B , 1 i K:. , where weights are calculated s are Eigenfaces. These as : . 248

I is chosen according to level of precision required in the results. xi

Volume 1

5 8 15 5

5 9 5 15

15 8 15 5

81.21 81.81 87.27 81.21

Iterations for subspace of dimension 2M/3 SVD (no. of vectors) 10 20 25 15 DFLDA (no. of vectors) 10 15 15 10 LPP (no. of vectors) 30 15 10 25 EFFICIENC Y (in %) 85.45 86.67 84.84 92.12

Figure 1

IV. EXPERIMENTAL RESULTS The method of choosing right combination of right proportion of feature vectors has been applied on a large database consisting of a variety of still images with illumination, expression variations as well as partially occluded images. The ratio 3:2:5::SVD::DF-LDA:LPP has yielded highest accuracy in recognition. The algorithm is implemented in MATLAB 7 on a Pentium Dual core 1GB RAM Windows XP workstation on a total test set of 165 images drawn from YALE dataset and the training set consisting 15 classes having a class count of five images. An ROC graph is plotted to visualize and analyze the working of face recognition efficiency. Iterations for subspace of dimension M/3 SVD (no. of vectors) DFLDA (no. of vectors) LPP (no. of vectors) EFFICIENC Y (in %)

SNAPSHOTS OF IMAGES

249
Volume 1

Training Set

Anecdotal experimentation with acquired image sets indicates that profile size, complexion, ambient lighting and facial angle play significant parts in the recognition of a particular image. Further research could be conducted into the viability of using nonlinear dimensionality reduction methods with neural network classifier. REFERENCES

Eigen faces from SVD

[1]

[2] [3] [4]

[5]

Fisher Faces from DF-LDA

[6]

[7]

Michael E. Wall, Andreas Rechtsteiner, Luis M. Rocha, Singular value decomposition and principal component analysis Modeling, Algorithms, and Informatics Group (CCS-3) Los Alamos National Laboratory, 1999 - 2009 M.A. Turk and A. Pentland, Eigenfaces for Recognition, Journal of Cognitive Neuro-science, vol. 3, pp.71-86, 1991. Xiaofei He,Shuicheng Yan, Yuxiao Hu*, Partha Niyogi, Hong-Jiang Zhang, Face Recognition Using Laplacianfaces, 2005 Juwei Lu, Kostantinos N. Plataniotis, and Anastasios N. Venetsanopoulos, Face Recognition Using LDA-Based Algorithms, IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 14, NO. 1, JANUARY 2003 Yu, H., Yang, J., A direct LDA algorithm for high-dimensional data - with application to face recognition. Pattern Recognition 34, 2067 2070, October 2001 Sam Roweis, and Lawrence K. Saul, Nonlinear Dimensionality Reduction by Locally Linear Embedding, Science, vol 290, 22 December 2000. Joshua B. Tenenbaum, Vin de Silva,John C. Langford, A Global Geometric Framework for Nonlinear Dimensionality Reduction, 7 August 2000; accepted 15 November 2000

Laplacian Faces from LPP

Query Images True Positives TN FP FN

V. CONCLUSION In this paper the approach to face recognition was motivated by information theory, leading to the idea of basing face recognition on a small set of image features that best approximate the set of known face images, without requiring that they correspond to our intuitive notions of facial parts and features. The approach does provide a practical solution to the problem of face recognition and is relatively simple and has been shown that it can work well in a constrained environment. 250
Volume 1