Focusing on the issue that the saliency detection algorithm via graph-based manifold ranking (MR algorithm) is over dependent on background features extracted from boundary nodes, an improved saliency detection algorithm combined with foreground and background features was proposed. Firstly, an image was divided into several super-pixels and a close-loop model was constructed. Secondly, the foreground and background seeds were obtained by using manifold ranking algorithm according to foreground and background features. Then these two kinds of seed nodes were combined through brightness and color characteristics, resulting in more accurate query nodes. Finally, a saliency map of the image was obtained by computing the saliency value via manifold ranking algorithm. Experimental results show that compared with MR algorithm, the precision rate, the recall rate and the F-measure of the proposed algorithm are significantly improved, and the obtained saliency maps are much more close to the true value.