Flexible robust principal component analysis

The error correction problem is a very important topic in machine learning. However, existing methods only focus on data recovery and ignore data compact representation. In this paper, we propose a flexible robust principal component analysis (FRPCA) method in which two different matrices are used to perform error correction and the data compact representation can be obtained by using one of matrices. Moreover, FRPCA selects the most relevant features to guarantee that the recovered data can faithfully preserve the original data semantics. The learning is done by solving a nuclear-norm regularized minimization problem, which is convex and can be solved in polynomial time. Experiments were conducted on image sequences containing targets of interest in a variety of environments, e.g., offices, campuses. We also compare our method with existing method in recovering the face images from corruptions. Experimental results show that the proposed method achieves better performances and it is more practical than the existing approaches.