Supervised extraction of face subspaces based on multimodal discriminant analysis

Abstract

Face image appearance may change due to a variety of factors (or modes) such as the person¿s identity, lighting condition and expression. We propose a method for representing these appearance changes as a mixture of di¿erent modes in di¿erent subspaces. These subspaces are simultaneously extracted in the following manner: we ¿rst transform the data to the whitened space and then perform Fisher Discriminant Analysis (FDA) to ¿nd mutually orthogonal discriminant subspaces for di¿erent modes based on their respective labeling information. The proposed method could be used for dimension reduction and face recognition. To validate the e¿ectiveness of the method, we have tested it on the Multi-PIE database. Experiment results of dimension reduction show satisfactory visual quality and those of face recognition show superior performance compared to PCA and LDA.