Saliency Detection by Compactness Diffusion

Most existing methods of salient object segmentation only focus on foreground cues such as contrast, or background cues such as boundary connectivity. Another problem is that they have used redundant information to generate an acceptable saliency map such as variances in different color spaces, multi-scale features and so on. In this paper, we propose saliency detecting with a diffusion model; use optimal seeds generated from foreground statistic cue, i.e., the compactness. Each superpixel is considered as a node and a fully connected graph is constructed to calculate the global compactness of each node. Then the local connected graph is constructed by only considering adjacent nodes, and compactness is diffused by applying a quadratic energy model to generate a coarse saliency map. After that, boundary prior is combined with the coarse saliency map for further eliminating the background. Experiments on three benchmark datasets including MSRA 1000, ECSSD and DUT-OMRON show that compared with other seven stateof-the-art methods, our model achieves stable and excellent performance. Parametric sensitivity analysis and time consumption are given to prove that the proposed method is stable and efficient.