Saliency Detection Based on Improved Manifold Ranking via Convex Hull

Aiming at the issue that traditional saliency detection algorithm via graph-based manifold ranking only relies on the boundary background prior saliency map to extract foreground seeds, affects the final ranking result, and makes saliency map worse, this paper proposes an algorithm that obtains more accurate foreground seeds to make manifold ranking by combining convex hull. First, a background saliency map is obtained by manifold ranking utilizing boundary nodes as background seeds, then a foreground region is got by a binary segmentation of the saliency map. Second, convex hull roughly containing saliency objects is constructed by Harris corner with color boosting of the input image. Finally, saliency map is computed by using manifold ranking algorithm, whose foreground seeds are from the combination of convex hull and foreground region. Experiment results show that the proposed algorithm gains improvement compared with some other classic algorithms in terms of PR curves, MAE values and F-measure on three public datasets.