Abstract: A novel two-dimensional maximum difference marginal Fisher discriminant analysis(2DMDMFA) was proposed for face recognition. The algorithm adopts the difference of similarity matrix Sp which characterizes the interclass reparability and similarity matrix S` which characterizes the intraclass compactness as discriminant criterion. In such a way,the small sample size problem occurred in marginal Fisher analysis(MFA) is avoided. In addition,the construction of Sp and Sr is directly based on original training image matrices rather vectors. It is not necessary to convert the image matrix into high-dimensional image vector like those previous methods so that the recognition rate is raised. Besides, the relations between the maximum difference marginal Fisher analysis discriminant criterion and marginal Fisher analysis discriminant criterion for feature extraction were revealed. Experimental results on ORL and Yale face database show that the algorithm outperforms the traditional methods in recognition performance.

Key words: Face recognition,Marginal Fishcr analysis (MFA)，Two-dimensional maximum difference marginal Fisher discriminant analysis(2DMDMFA)，Image matrix