Robust saliency propagation based on random walks

Abstract

A variety saliency detection methods have been based on the background prior knowledge. Nevertheless, some image patches connected to the image boundary are foreground noise. In this paper, we propose a saliency detection algorithm which propagates coarse saliency based on robust background prior set via random walks. First, we get the robust background prior set from the initial background set which is made up by all image patches connected to the image boundary, and compute a coarse saliency map through summation of global contrast between image patch in complement set of robust background prior set and image patch in robust background set. Second, we get foreground prior set by segmenting the coarse saliency via adaptive threshold, and propagate foreground prior set via random walks. Propagated foreground prior set is integrated by Bayesian inference model and resulting a saliency map. Third, we update saliency map by hypergraph and integrate current saliency and updated saliency by weighted mean summation. Finally, results from multi-scale saliency maps are integrated by pixel-wise weighted summation. Experimental result shows that our approach outperforms state-of-art approaches on two public available dataset.