Face recognition with continuous occlusion using partially iteratively reweighted sparse coding

Partially occluded faces are common in automatic face recognition in the real world. Existing methods, such as sparse error correction with Markov random fields, correntropy-based sparse representation and robust sparse coding, are all based on error correction, which relies on the perfect reconstruction of the occluded facial image and limits their recognition rates especially when the occluded regions are large. It helps to enhance recognition rates if we can detect the occluded portions and exclude them from further classification. Based on a magnitude order measure, we propose an innovative effective occlusion detection algorithm, called Partially Iteratively Reweighted Sparse Coding (PIRSC). Compared to the state-of-the-art methods, our PIRSC based classifier greatly improve the face recognition rate especially when the occlusion percentage is large.