Null Space Marginal Fisher Analysis and Its Application in Face Recognition

Marginal Fisher analysis (MFA) is an efficient linear projection technique for feature extraction. The major drawback of applying MFA to face recognition is that it often encounters the small sample size (SSS) problem. In this paper, a strategy based on null space for solving optimization criteria of MFA is proposed to avoid this issue. It maximizes the class scatter of training samples on null space of within-class scatter matrix (S_w) in MFA and reserves the discriminant information contained in null space of S_w. The performance of this method is tested in both ORL and Yale face databases. Experimental results show that this method is effective and achieves higher recognition rate than LDA and MFA. Moreover, it is easy to decide most optimal dimensionality of feature space for this method.