Multi-scale self-searching saliency detection combined with rectangular diffusion

According to the characteristics of human visual system, a multi-scale self-search saliency detection method combined with the rectangular diffusion is proposed. In this paper, we first measure saliency from foreground and background respectively and then fuse them. In foreground view, we rank superpixels on center point, inner layer rectangle, middle layer rectangle and outer layer rectangle in the image as seed nodes by manifold ranking, which is named rectangular diffusion, and then fuse them. Outer layer rectangle diffusion need to be corrected for it includes little saliency object. In background view, we first sort all the superpixels by their feature weight of CIELab color space, and then drop the least 60% superpixels, which is named self-searching, and last remove superpixels on middle layer rectangle to form background seed point, and then apply manifold ranking to generate saliency map from background view. In comparison experiments on datasets of MSRA-1000, CSSD, ECSSD, THUS-10000 and SED, our method performs well against the state-of-the-art methods. It gets highest precision-recall curve on five datasets, which fully shows its effectiveness.