Abstract—In face recognition, feature extraction techniques attempts to search for appropriate representation of the data. However, when the feature dimension is larger than the samples size, it brings performance degradation. Hence, we propose a method called Normalization Discriminant Independent Component Analysis (NDICA). The input data will be regularized to obtain the most reliable features from the data and processed using Independent Component Analysis (ICA). The proposed method is evaluated on three face databases, Olivetti Research Ltd (ORL), Face Recognition Technology (FERET) and Face Recognition Grand Challenge (FRGC). NDICA showed it effectiveness compared with other unsupervised and supervised techniques. Keywords—Face recognition, small sample size, regularization, independent component analysis. I. INTRODUCTION VER the past decades, biometrics authentication technology has been widely developed for security purpose. Among these technologies, face recognition has gained considerable attention in recent years. There are numerous applications of face recognition such as in government use and commercial use, e.g. access control, surveillance purpose, banking etc. Compared with other biometrics technologies such as iris authentication, finger authentication and etc, there are some reasons for its extensive attention. Face recognition is an authentication system which does not require expensive and advance devices. Furthermore, it is contactless as the authentication process can be done without voluntary action from the users. Normally, illumination, pose, lighting and facial expression are the variants that result system performance degradation. Many approaches have been applied on the face recognition for better improvement [1]-[5]. Generally, face analysis approaches could be classified into three categories such as holistic approach, feature-based approach and hybrid approach. Holistic approach makes use of the global information which derived from facial image pixels. The representative examples of holistic approach are Principal Component Analysis (PCA) by Kirby and Sirovich (1988) [6] and Linear Discriminant Analysis (LDA) which proposed by Peter N. Belhumeur et al. in 1997 [7]. Feature-based approach adopts local features of the face for data learning. These local features include eyes, nose, This work is supported by the financial support of Telekom Research and Development Sdn Bhd. of Malaysia. The authors are with the Faculty of Information Science and Technology, Multimedia University, 75450 Melaka, Malaysia. (email: lyping8@yahoo.com, yhpang@mmu.edu.my, lau.siong.hoe@mmu.edu.my, syooi@mmu.edu.my, me.the.fren@gmail.com). mouth, chin and head line. Elastic Bunch Graph Matching (EBGM) is one of the famous feature-based approaches which proposed by Wiskott, L et al. (1997) in face recognition [5]. Hybrid approach utilizes both holistic and feature-based approaches. The idea of this hybrid approach is that human facial shares the same basic features such as nose, mouth but challenging in distinguish the characteristics which are not present in other face component such as forehead and chin. Neither holistic nor feature-based approach could analyse these two features at a time. Hence, hybrid approach has the capability to perform this computation. One of the face recognition algorithm based on hybrid approach is proposed by Shiladitya Chowdhury et al. (2010) [8]. They were using generalized two-dimensional Fisher’s Linear Discriminant method. The method adopts maximization of class separability from both the row and column directions simultaneously and yields much smaller image feature matrix. PCA is one of the popular appearance-based techniques in the face recognition. PCA is also known as Eigenfaces. By using eigenspace decomposition of covariance matrix which derived from the training sample, PCA approximates linearly independent basis of face subspace. In PCA analysis, the whole face image is taken as a whole. This leads to the image variation from the same person (within-class scatter) could be larger than the image variation of face identity (between-class scatter). Usually, this within-class variation is induced by varying illumination, pose or facial expression of the individual. Based on this observation, PCA is enhanced through adopting discriminant criterion. This enhanced technique is known as LDA. LDA defines a projection which maximizes the between-class scatter and minimize within-class scatter. However, the decorrelation of input data is based on second-order statistics and ignoring higher order information. In face recognition, it is believed that there are some significant information contained in the higher-order relationship of image pixels. Therefore, ICA is proposed. Independent Component Analysis (ICA) is another technique used in face recognition. ICA is a technique to separate independent component from a mixed source. It uses high order statistics as presented in [3]; ICA has two architectures, Architecture I and Architecture II. ICA Architecture I is an algorithm where the images are treated as variables and pixels are the observations. It focused on spatially localized features. For ICA Architecture II, it treats images as observation and pixels as variable. It is mainly focused on global features. There are several extensive researches on ICA face recognition [9]-[11]. Normalization Discriminant Independent Component Analysis Liew Yee Ping, Pang Ying Han, Lau Siong Hoe, Ooi Shih Yin, and Housam Khalifa Bashier Babiker O World Academy of Science, Engineering and Technology International Journal of Computer and Information Engineering Vol:7, No:8, 2013 1099 International Scholarly and Scientific Research & Innovation 7(8) 2013 scholar.waset.org/1307-6892/16147 International Science Index, Computer and Information Engineering Vol:7, No:8, 2013 waset.org/Publication/16147