Face recognition using FLDA with single training image per person

Keywords: Face recognition Fisher linear discriminant analysis (FLDA) Single training image per person Singular value decomposition (SVD) a b s t r a c t Fisher linear discriminant analysis (FLDA) has been widely used for feature extraction in face recognition. However, it cannot be used when each object has only one training sample because the intra-class variations cannot be statistically measured in this case. In this paper, a novel method is proposed to solve this problem by evaluating the within-class scatter matrix from the available single training image. By using singular value decomposition (SVD), we decompose the face image into two complementary parts: a smooth general appearance image and a difference image. The later is used to approximately evaluate the within-class scatter matrix and thus the FLDA can be applied to extract the discrimi-nant face features. Experimental results show that the proposed method is efficient and it can achieve higher recognition accuracy than many existing schemes. Face recognition has been extensively studied for many years and it is still attracting much attention because of its big potential in security, surveillance and human-computer intelligent interaction, etc. A key issue in face recognition is to find sufficient and discriminative features for face representation. Many approaches have been proposed and subspace analysis method (SAM) has become one of the most popular methods. SAM seeks for a set of basis vectors according to some criteria and extracts the features by projecting face images onto the subspace spanned by those basis vectors [1–3]. Principal component analysis (PCA), which tries to find a set of optimal orthogonal bases in the sense of minimum mean square error, and Fisher linear discriminant analysis (FLDA), which tries to find a set of optimal projection vectors by maximizing the ratio between the determinants of the between-class and the within-class scatter matrices of the training samples , are the two most representative methods in SAM. By first applying PCA to face recognition, Kirby and Sirovich [4] found that a face image could be reconstructed approximately as a weighted sum of a small collection of basis face images plus a mean face image. Based on this work, Turk and Pentland [5] developed the well-known Eigenface method. Since then, PCA has been extensively investigated and many PCA-based algorithms have been developed [6,7]. Although PCA enables sufficient reconstruction, it may not be optimal for classification because its optimality is in the sense of minimum …