The traditional random walks based image segmentation algorithm requires setting seed points interactively to obtain the desired segmentation results. Based on visual attention, the paper proposes a new random walks based image segmentation algorithm with the seed points determined automatically. Firstly, the probability boundary map (PBM) is generated and the image is divided into super pixels. Then, the key segmentation region is searched by shifting visual attention focus with Itti model. In order to determine the seed points of the key segmentation region, the probabilistic boundary map is transformed into polar coordinates map taking the current focus of attention as the pole. The energy function about the boundary of the focal region is established on the obtained polar coordinate probabilistic boundary map. The energy function can be minimized by the maxflow mincut algorithm, and the super pixels within the boundary of focal region are the seed points of segmentation region. Finally, super pixels of images are used as nodes to construct a graph, random walks algorithm is conducted on the graph to complete the image segmentation. The experiments on Berkeley Segmentation Data Set show that the proposed method is effective to complex images’ segmentation.