Similarity mapping for robust face recognition via a single training sample per person

Abstract

Face recognition using a single training sample per person (STSPP) is very challenging due to significant variations in the query sample. Although most of sparse representation based methods eliminate the variations to obtain discriminative power, their recognition performance usually degrades as the number of training samples per person decreases. To improve recognition performance on the STSPP problem, in this paper, we propose a novel face recognition method: generating new training samples by preserving the variations of each query sample onto training samples instead of eliminating the variations, and then identifying the query one via the most similar generated training sample. Specifically, we formulate face recognition as an patch-based sparse representation problem, and then efficiently solve it by dividing the optimization procedure into three individual sub-problems: selecting training patches, weighting selected patches and calculating residuals. Meanwhile, we interpret the proposed algorithm from a probability modeling viewpoint. Experiments on multiple benchmark databases demonstrate its superior recognition performance to state-of-the-art methods on faces under illumination changes, occlusions and facial expressions, with lower test time costs.