Robust Face Recognition Based on Spatially-Weighted Sparse Coding

Peipei Zhang, Huicheng Zheng, Chun Yang

Abstract: Recently sparse representation has been widely used in face recognition. It has been shown that under maximum likelihood estimation of the sparse coding problem, robustness of face representation and recognition can be improved. In this paper, we propose to weight spatial locations based on their discriminabilities in sparse coding for robust face recognition. More specifically, we estimate the weights at image locations based on a class-specific discriminative scheme, so as to highlight locations in face images that are important for classification. Furthermore, since neighboring locations in face images are often strongly correlated, spatial weights are smoothed to enforce similar values at adjacent locations. Extensive experiments on benchmark face databases demonstrate that our method is very effective in dealing with face occlusion, corruption, lighting and expression changes, etc.