I. INTRODUCTION
Feature extraction is the key to face recognition. The aim of feature extraction is to reduce the dimensionality of face image so that the extracted features are as representative as possible. When dealing with images, we should firstly transform the image matrixes into image vectors. Then based on these vectors, the between-class scatter matrix, within-class scatter matrix and total-class scatter matrix are calculated and the optimal projection axes are obtained. However, face image recognition is of high dimension and small sample size problem. The traditional FDA encounters two aspects of difficulties. First, the traditional algorithm cannot be used directly when the within-class scatter matrix is singular. Second, the high dimensional image vectors lead to computational difficulty. To overcome the drawback, Yang et al. [1] proposed a straightforward image projection technique, named image principal component analysis (i.e. 2DPCA); the method is to exploit image matrices to directly construct between class scatter matrix and within class scatter matrix.