Graph construction by incorporating local and global affinity graphs for saliency detection

The diffusion-based graph is one type of extensively studied method in saliency detection due to its simplicity and effectiveness. The key to success is a superior affinity graph and suitable initial seeds. In this paper, considering the local structure and global confidence of the affinity graph, we generate a joint global affinity graph from the deep-level and low-level image feature matrix based on the self-representation model, as well as a deep-level local affinity graph and a low-level local affinity graph using these two types of features, respectively. Based on the local and global affinity graph, we design three superior affinity graphs to benefit the corresponding three stages consisting of saliency estimation, saliency refinement, and saliency optimization. In particular, in the second and third stages, we embed the initial saliency values to robust the affinity graphs. Extensive experiments demonstrate that the proposed method over performing the state-of-the-art approaches on four benchmark datasets.