Kernel Fisher Discriminant Analysis in Full Eigenspace

This work proposes a method which enables us to perform kernel Fisher discriminant analysis in the whole eigenspace for face recognition. It employs the ratio of eigenvalues to decompose the entire kernel feature space into two subspaces: a reliable subspace spanned mainly by the facial variation and an unreliable subspace due to finite number of training samples. Eigenvectors are then scaled using a suitable weighting function. This weighting function circumvents undue scaling of projection vectors corresponding to the undependable small and zero eigenvalues. Eigenfeatures are only extracted after the discriminant evaluation in the whole kernel feature space. These efforts facilitate a discriminative and stable low-dimensional feature representation of the face image. Experimental results comparing other popular kernel subspace methods on FERET, ORL and GT databases show that our approach consistently outperforms others.