Optimized multi-task sparse representation based classification method for robust face recognition

Sparse representation based classification (SRC) method has become a hot topic in recent years. To address the alignment problem, feature-based (such as SIFT) SRC method has been proposed. But it always works not ideally for the contiguously occluded face images. After analyzing, we point it to the issue of features´ reliability for multi-task recognition. Through the theoretical analysis on the sparsity of SR coefficient, a formula for evaluating features´ representation reliability (RR) is proposed to optimize the multi-task feature-based SRC. In this paper, firstly, we present the main thought of the proposed formula of RR. Then it is introduced to the feature-based SRC method. Finally, experiments on Yale and AR database are performed. The proposed method is compared with the methods of MKD-SRC, SIFT-matching and original SRC. Experimental results show that the proposed method is more robust for simultaneously handling the variations in illumination, alignment, expression, and occlusion.