A global and local consistent ranking model for image saliency computation

Image saliency detection is an important issue in computer vision and has been widely used in many applications. In this paper, we propose a new global and local consistent ranking (GLR) model for image saliency computation. Firstly, we propose to use an absorbed Markov chain model to obtain a kind of global ranking for image superpixels, in which the absorbing nodes represent the virtual boundary superpixels and the transient nodes denote the general superpixels of image. Then, the absorbed time from each transient node to boundary absorbing nodes are computed. This absorbing time of transient node measures its global similarity with all absorbing nodes and thus provides a kind of global ranking for each transient node w.r.t. absorbing nodes. At last, we further exploit the local manifold structure and incorporate the local manifold smooth information into ranking process and thus propose a general global and local consistent ranking for saliency detection. Experimental results on several large benchmark databases show the effectiveness of the proposed GLR method.