A Multiple Graph Label Propagation Integration Framework for Salient Object Detection

Saliency prediction typically relies on multiple features that are combined in the ways of weighted summation or multiplication to form a saliency map, which is heuristic and hard for generalization. In this paper, a novel multiple graph label propagation integration framework for saliency object detection algorithm is proposed. The proposed algorithm is divided into four steps. First, an input image is segmented into superpixels which are represented as nodes in a graph and transformed from RGB color space into CIE L*a*b* color space. Second, combined by texture features, we measure the similarity of two adjacent superpixels for each feature, which is represented as an affinity matrix. Then, to generate the salient seeds, we adopt the color boosting Harris points as salient points to catch the corners or marginal points of visual salient region in color image. The saliency points provide us a coarse location of the salient areas. In the last step, the graphs are combined into label propagation framework to obtain the saliency maps. We propose efficient optimization algorithms for the proposed approach, which generate sparse weighted coefficients that allow identifying the graphs which are important or not for salient object detection easily. Experiments on four benchmark databases demonstrate the proposed method performs well when it violates the state-of-the-art methods in terms of accuracy and robustness.