Salient Object Detection based on Multiple Priors Fusion

Abstract

Salient object detection can be utilized to detect the most significant regions in various environments, which has been regarded as foundation of computer vision. Different saliency models use different prior or knowledge. We propose a multi priors fusion method for saliency measure, which integrates background prior with foreground prior and center prior. Firstly, through each boundary of the image, we can get four saliency maps, and fuse them to get the background prior saliency map. Secondly, we utilize boundary extension method to highlight regions, and these regions can be regarded as the queries of manifold ranking for the foreground prior saliency map. Thirdly, the corners on the image are obtained, filtered by the foreground region, and then clustered into a point as the center of Gaussian model, which is used to calculate the center prior saliency map. Finally, the above three kinds of prior-based saliency maps are fused via the proposed fusion framework to gain a better final saliency map. Compared with fifteen methods, the experimental results on ECSSD and MSRA10K show that our proposed method achieves better saliency detection results.