Robust Face Recognition Based l21-Norm Sparse Representation

In recent years, Sparse Representation based classification (SRC) has made great progress in Face Recognition. However, SRC is only efficient and effective when the noise is sparse. The recognition rate of SRC decreases when the noise is non-Gaussian, for example, the light on the face is quite various or the face is covered in part by a mask. In this paper, we propose a robust l2,1-norm Sparse Representation framework that constrains the noise penalty by the l2,1-norm. This framework takes both advantages of the discriminative nature of the l1-norm and the systemic representation of the l2-norm. In addition, we also use the l-norm to constrain the coefficient matrix. As the l-norm concerns the global structure, our method is robust to the noise, especially for the case when the contiguous occlusion exists in the real world. The extensive experiments demonstrate that when dealing with large region contiguous occlusion, the proposed method achieves significantly better results than SRC and some other sparse representation based face recognition methods.