Saliency Detection Based on Non-convex Weighted Surrogates

Low rank and sparsity decomposition have shown potential for salient object detection. In existing methods, nuclear norm is used to approximate rank minimization and 1(l) norm is selected as sparse regularization. Two deficiencies, however, still exist for nuclear norm and 1(l) norm. First, both always over-penalize large singular values or large entries of vectors and result in a biased solution. Second, the existing algorithms very slow for large-scale applications. To address these problems, we propose a novel weighted matrix decomposition model with two regularizations: (1) Schatten-2/3 quasi-norm that captures the lower rank of background by matrix factorization technique, and (2) The l(2/3) - norm that is capable of producing consistent salient object within the same image patches by effectively absorbing both image geometrical structure and feature similarity. In addition, we equip the weighting matrix with a high-level background prior map based on the color, location and boundary connectivity, which can indicate the probability that each image region belongs to the background. The proposed model can be solved by perform SVDs on two much smaller factor matrices. Experiments on three broadly used datasets by detailed comparisons show that our proposed approach has potential in salient object detection.