Saliency Detection Based on Robust Foreground Selection

Saliency detection is to find the most important object automatically according to the human visual in the unknown scene. For improving the precision of saliency detection, the saliency detection based on robust foreground seeds via manifold ranking is proposed in this paper. Firstly, the two different convex hulls are got by the Harris corner and boundary connectivity algorithm. And the original object region is defined by the intersection about the above convex hulls. Secondly, the superpixels in convex hull are done the similarity detection with the outer edge of the convex hull. The superpixels are removed when they are similar to most of the outer edge, and the more precision foreground seeds are got. Using the anchor graph, a novel graph construction is built to express the relationship between data nodes. And then, two different kinds of salient results will be got based on ranking on manifolds using foreground and background seeds respectively. Finally, the saliency map is got through optimizing a novel cost function. Experimental results prove that the proposed algorithm improves the performance evaluation of precision and recall rate further.