Improved saliency detection based on manifold ranking algorithm
Liang Yao, Hongliang Chen, Jianxun Li · 2017
Saliency detection is a fundamental problem in computational and cognitive sciences. Nowadays, graph-based methods are widely applied to saliency detection including manifold ranking(MR) method, which is shown to be fast and effective. However, because of only using a single feature and imperfect selection strategy for background seeds, MR has a poor performance in some circumstances. In order to complete more challenging saliency detection, an improved method based on manifold ranking algorithm is proposed in this paper. The adopted parallel architecture enables the final detection results generated with the combination of calculations from background seeds and foreground seeds. These seeds are quickly obtained by using priori information and the graph structure is optimized with adding three constraints. Moreover, the graph edge weights are computed by utilizing a adaptive local width parameter and measuring multi-features distance. Three-stage strategy is used to calculate the final saliency map. Experiment results on two large famous datasets demonstrate that the proposed method performs better comparing with MR and other state-of-the-art methods.