Three-Stage Salient Object Detection based on Integrated Priors

Abstract

In this paper, a three-stage model is proposed for salient object detection. In the proposed method, an intuitive and straightforward pre-treatment method is firstly proposed to conduct superpixel segmentation adaptively, then superpixel-based graphs are constructed to express the structure of the image. To make full use of the information of individual images, multiple priors, including background prior, foreground prior, center prior and global contrast prior, are integrated in the three-stage detection model. In the first stage, under the assumption of background prior that the borders of the image are more likely to be the background, the absorbing Markov chain model is constructed to compute the saliency scores based on the absorbed time of each node in random walk. Then in the second stage, the saliency scores computed in the first stage, are taken as the foreground prior to compute the saliency scores via manifold ranking. In the third stage, center-biased global contrast filter combining center prior and global contrast prior is formulated to refine the saliency map. Experimental results demonstrate the effectiveness of the proposed three-stage method.