Salient object detection based on Laplacian similarity metrics

Background prior has become a novel viewpoint and made progresses in salient object detection. Most existing salient object detection algorithms based on background prior take boundaries as backgrounds and neglect nonbackground factors of boundaries, which is in fact unreasonable. Thus it is necessary to combine background prior with the analysis of boundary property. In this paper, the probability values computed by Mahalanobis distance are used to describe the likelihood of boundary superpixels belonging to backgrounds, which is viewed as a method for analyzing boundary properties. Meanwhile, some cues should be integrated with the obtained probability values for saliency computation. Inspired by the theory of Laplacian similarity metrics, two-stage complementary metrics are established according to different clusters in which two-stage queries lie, and a two-stage detection algorithm (SLSM) of salient objects is thus proposed by combining two-stage complementary similarity metrics with the probability values. Furthermore, when the detailed clusters (dense or sparse) of queries in each detection stage are ignored, an additional unified similarity metric is also constructed. Through the combination of the unified similarity metric and the proposed method for analyzing the boundary properties, another baseline algorithm (SLSMU) is also created. The results of experiments in which these two proposed algorithms are applied to four datasets demonstrate each of the two algorithms outperforms some existing state-of-the-art methods in terms of the different metrics.