Saliency map fusion based on rank-one constraint

Abstract

Co-saliency is the common saliency existing in multiple images, which keeps consistent in saliency maps. One saliency detection method generates saliency maps for all the input images, so that we have a group of maps. Salient region of each image is extracted by its corresponding saliency map in the group. We use a matrix to combine all the salient regions. Ideally, these co-salient regions are similar and consistent, and therefore the matrix rank appears low. In this paper, we formalize this general consistency criterion as rankone constraint and propose a consistency energy to measure the approximation degree between matrix rank and one. We combine the single and multiple image saliency maps, and adaptively weight these maps under the rank-one constraint to generate the co-saliency map. Our method is valid for more than two input images and has more robustness than the existing co-saliency methods. Experimental results on benchmark database demonstrate that our method has the satisfactory performance on co-saliency detection.