Abstract :
In this paper, a nonconvex regularization based matrix decomposition model (NRMD) is proposed. In NRMD, the non-salient regions are regarded as a low rank part, and the salient regions are viewed as a sparse part. Different from previous methods, the non-salient regions are constrained by Sp-norm (0 <; p <; 1). Sp-norm with an appropriate choice of value p may force the background more low rank. In addition, a group sparsity induced norm, which is imposed on salient regions, is introduced to describe potential spatial relations of patches and that of the groups. The Frobenius norm guarantees patches sharing similar representations within the same group. Thus, patches in s same group may get similar saliency values. The optimization problem can be solved through an augmented Lagrange multipliers method. Finally, high-level priors are fused to the matrix division and boom the salient regions detection. Experiments on the widely used MRSA-1000 dataset show that NRMD outperforms the state-of-the-art models.
