Salient Object Detection Integrating Both Background and Foreground Information Based on Manifold Preserving

Graph-based two-stage algorithms have widely developed and achieved good performance to detect salient objects. For these algorithms, choosing the proper seeds using for saliency propagation is quite crucial and difficult. In this paper, we consider using background/foreground probability values of candidate background/foreground seeds as the estimation of the reliable seeds, not considering the refinement of candidate seeds. Moreover, these probability values are integrated into the proposed saliency models, which can avoid hard filtering for candidate seeds as well as simplify the procedure of the algorithm. In addition, considering the manifold structure of an image, we fuse the manifold-preserving term into the saliency models. Especially, reconstruction matrix $A$ is determined based on the deep features extracted from FCN-32s, which can further improve detection performance of salient objects. The results of experiments in which the proposed SBFMP algorithm is applied to four datasets demonstrate SBFMP algorithm is prior to some existing state-of-the-art algorithms in terms of the different metrics.