Contrast based Saliency Detection via Manifold Ranking

Abstract

Abstract. Saliency detection has been widely studied in computer vision. In this paper we propose a two-steps method combining contrast assumption and ranking technology to detect saliency region. Firstly, We use the traditional contrast assumption to find foregroundcues. Then we rank the nodes with labeled contrast cues bygraph-based manifold ranking. We experiment with our method on a large public data set. Our results show the effectiveness of our method, and perform better compared to recent state-of-the-art methods. INTRODUCTIONIn computer vision, salient region detection is a long-standing problem. It measures the quality of the region standing out to its neighbors. Usually, salient regions are those attracting the most attention when observers glance the scene. It is a complex problemidentifying these regionswithout prior knowledge, and it lays in the interdisciplinary of computer science,biology, psychology and other fields. Hundreds of computational saliency models that following early attention models (e.g., Koch and Ullman