Saliency Detection: Multi-Level Combination Approach via Graph-Based Manifold Ranking

In this paper, we form the saliency map by a multi-level combination approach using graph-based manifold ranking method as the foundation. This method aims to overcome the defects that the merely graph-based manifold ranking method can not do well in highlighting the salient objects uniformly and suppressing the background effectively. We first execute over-segmentation algorithm for each input image. In order to take full advantages of the image cues, input image is segmented into multilevels from the coarsest to the finest in terms of the number of segmented regions. And for each level, graph-based manifold ranking approach is used to produce the initial saliency map. Then we obtain the optimal weight of each level from the trained combiner to make a multi-level combination to obtain the final saliency map. The combiner is trained using quadratic programming algorithm on the MSRA-10K database. In the experiment, we compare our method with the other eight state-of-the-art methods according to four evaluation criteria. The experimental results show that the proposed algorithm achieves favorable performance compared with the other eight state-of-the-art approaches on different databases.