Maximal Margin Local Preserving Median Fisher Discriminant Analysis for Face Recognition

Median Fisher Discriminator(MFD) used the class median vector is more effective than Linear discriminant analysis(LDA). However, MFD only captures global geometrical structure information of the data and ignores the geometrical structure information of local data point. In this paper, we introduce a linear approach, called Maximal Margin Local Preserving Median Fisher Discriminant Analysis (MMLPMFDA). MMLPMFDA models the geometrical structure and variability of the local neighborhoods by constructing two adjacency graphs over the training data, and then incorporates the geometry and variability into the objective function of the MFD. In order to solve the small sample size problem, the objective function in a form of the difference is adopted. Finally, experiments on the ORL, YALE and AR face databases show the effectiveness of the proposed approach.