Similar Region Contrast Based Salient Object Detection

Abstract

Detection of visual saliency is an important issue in many computer vision tasks. In this paper, we propose a novel regional contrast based saliency detection method, generating a saliency map that enables high contrast between the foreground salient object and background. Our method mainly integrates four principles, which are based on psychological evidences, visual research and general observation. In order to suppress the homogeneous regions, and let the novel regions stand out, our method computes a region’s saliency value based on the region’s N closest regions defined in the CIE L*a*b color space. We compared our method with the state-of-the-art saliency detection methods using a standard publicly available database. Experimental results show that our method has better performance on yielding higher precision and recall rates. In the application of image editing, we demonstrate that using our saliency map as energy map can achieve more appealing retargeting results with less distortions in the important regions.