Linear generalization probe samples for face recognition

Abstract

In face recognition system, achieving higher recognition rate by limited training images is a challenging problem. An idea to solve this problem is generalizing some virtual probes from original probes such as nearest feature line (NFL) [IEEE Trans. Neural Networks 10 (1999) 439], nearest feature angle (NFA) and simple hybrid classier (SHC) (hybrid NFA and NFL) method [Pattern Recognition Lett. 23 (2002) 833].In this paper, an improved method for generalizing probe sets called linear generalization subspace (LGS) is proposed, in which the generalized area is some constrained linear subspaces of the original probes. In LGS, a method called constraint least square residual distance is suggested in this paper. Experimental results show that the proposed method has lower recognition error rate than the nearest neighbor method (without generalization), NFL and SHC, respectively, and low computation complexity than NFL or SHC methods based on ORL face database and a larger combination face database.