Single-Image Expression Invariant Face Recognition Based on Sparse Representation

Face recognition under expression variation has been paid few attentions and remains a difficult problem. This is because the non-linear shape variation makes it infeasible to match two images linearly. This paper believes that expression variations affects the face recognition in three aspects, i.e., alignment error, shape change, and occlusion. Based on this observation, this paper solves this problem using a shape-constrained sparse representation (SSR) framework. The proposed method presumes that expression variations have dramatic effect on shape, along with minor effect on texture such as occlusion. It has two contributions. First, SSR introduces a shape-constrained texture matching (STM) algorithm to solve the alignment error and the shape changes. This strategy is able to eliminate shape changes as long as the two objects have similar textures. This is different from state-of-the-art algorithms which directly match two faces using learnt measures beforehand. Second, SSR matches the two aligned images based on the sparse representation theory. As a result, an image can be sparsely represented by the image with the same identity, even if the texture has been partly occluded. Extensive experiments show that SSR obtains the state-of-the-art performance robustly.