Improved Bayesian Saliency Detection Based on BING and Graph Model

ABSTRACT

Saliency detection plays an important role in many computer vision applications. The traditional Bayesian based saliency model is used based on convex hull to circle a coarse salient region, which is inaccurate and unstable. To address this problem, this paper proposed an improved Bayesian framework based saliency method. Firstly, the BING (Binarized Normed Gradients) method was utilized to generate the coarse conspicuity map. Then, a graph model was constructed after SLIC superpixel image abstraction, to refine the initial conspicuity map. This is followed by the spatial information based weighting, to produce the final prior map. Secondly, after adaptive threshold, the observation likelihood map was computed by color histogram. Finally, these two maps were combined using Bayesian formula. Experimental results on two benchmark datasets MSRA-1000 and SOD show that the improved method was superior to 13 state-of-the-art alternatives, especially the previous Bayesian saliency models.

Keywords: