Abstract
Visual saliency detection (VSD) has been attracting increasing attention due to its wide applications in computer visions. In this paper, a visual saliency detection method based on maximum entropy random walk (MEVSD) is proposed. Gaze wandering over images is modeled as a random walk process on a graph, in which the super-pixels and their similarities are regarded as nodes and edges respectively. The visual salient region is detected based on the stationary distribution of the stochastic process. Compared with the stat-of-the-art methods, MEVSD focuses on globally maximizing the entropy of walking trajectories. A two-tier neighbor scheme of super-pixels is proposed to reduce the computational complexity. Besides, a belief diffusion algorithm on weight matrix is proposed to improve precision rate. Extensive experiments with diverse content of pictures indicate that MEVSD performs well compared with 11 state-of-the-art methods.