EMR: Extended Manifold Ranking for Saliency Detection.

A novel and outstanding saliency detection approach based on color features and background prior is proposed in this paper. Specifically, background prior is used in saliency detection widely, which considers the image boundaries as part of background. Then we propose an extended manifold ranking (EMR) algorithm to propagate the background prior to other image regions. Compared with GMR, EMR eliminates the negative effect of the initial assumption that non-boundary areas are all saliency regions. Furthermore, gradient boosting decision tree (GBDT) is introduced to refine the saliency map generated by EMR. The experimental results on three benchmark datasets demonstrate that our algorithm outperforms 10 state-of-the-art methods based on low-level features.