Manifold Smoothed Class-specific Discriminant Collaborative Representation for Face Recognition

Abstract

Representation based classifications are extensively used in many applications and have achieved remarkable successes. However, they still suffer from some problems. In one respect, the coefficients obtained are not competitive, which is not beneficial for classification. Also, the intrinsic manifold structure is not taken into consideration. It has been shown that the manifold smoothness is important for classification. To this end, based on collaborative representation, a new algorithm coined Manifold Smoothed Class-specific Discriminant Collaborative Representation, MSCDCR, for short, is proposed. Besides manifold structure, MSCDCR employs the whole training samples to code the test sample in a linear way, so that it can tell which class has the most similarity. The method proposed can reduced to a closed-form solution, which is time-efficient, and it also can exploit the intrinsic manifold geometric structure and discriminant information. Consequently, it bears more discriminant ability. Experiments are carried out on face recognition, and they show the relatively promising performance of the proposed algorithm.