Double Discriminant Analysis for Face Recognition

Feature selection for face representation is one of the central issues for any face recognition system. Finding a lower dimensional feature space with enhanced discriminating power is one of the important tasks. The traditional subspace methods represent each face image as a point in the disciminant subspace that is shared by all faces of different subject (classes). Such type of representation fails to accurately represent the most discriminate features related to one class of face, so in order to extract features that capture a particular class’s notion of similarity and differ much from remaining classes is modeled. In this paper we propose a new method called “Double Discriminant Analysis” Which first performs PCA (Principal Component Analysis) to reduce the sample size and extract the features that separates individual class faces maximally. Then by projecting these samples over to the null space of within class matrix the intra class variance is reduced to extract the most discriminative feature vectors for which an Individual class oriented subspace is found for each class ‘i’ along which the intra class variance is minimum, and separates well from all the remaining classes. This individual subspace for each class ‘i’ found to express most discriminative power that helps in classification and thus developing an effective face recognition system.