Due to recent rapid development of computer vision applications such as object recognition and image segmentation, it has become increasingly important to generate reliable saliency maps to uniformly highlight the desired salient object. In this paper, we present a novel bottom-up salient region detection method by exploiting contrast prior and the relationship between the salient region detection and graph based semi-supervised learning problem. First, we compute a preliminary initial saliency map by a newly proposed technique named unit boundary distribution and several refinement schemes. Second, after obtaining the indication map generated via a double threshold operation on the initial saliency map, we model the final saliency inference problem as a graph based semi-supervised learning approach by solving a energy minimization problem. Both quantitative and qualitative evaluations on three widely used datasets demonstrate the superiority of the proposed method to other twenty-one state-of-the-art methods.