Feature selection based saliency object detection

Color is essential feature in computer vision, to find a distinct color representation of the foreground and background is difficult. In this paper, we propose a novel method to pursue color features which are distinguishable for foreground and background. To achieve the initial position of the foreground, we impose the bi-segmentation mask of saliency map. However, single saliency map could not ensure the quality of the initialization. Such that we use the mask of the bi-segment the average of different saliency maps as initial seed. The distinct color features are selected by our feature selection method based on the foreground and background mask. Then we build a graph on the super-pixel segmentation, and the affinity matrix is computed based on the combined features. The new features endow higher similarity to the edges in the foreground (or background), but endow lower similarity to the edges across the foreground and background. Then we impose manifold ranking method to compute the final saliency maps. Our systematical experimental evaluations show that the proposed method can produce competitive results in comparison to the state-of-the-art.