Saliency Detection Inspired by Topological Perception Theory

The topological perception theory claims that visual perception of a scene begins from topological properties and then exploits local details. Inspired by this theory, we defined the topological descriptor and topological complexity, and we observed, based on statistics, that the saliencies of the regions with higher topological complexities are generally higher than those of regions with lower topological complexities. We then introduced the topological complexity as a saliency prior and proposed a novel unsupervised topo-prior-guided saliency detection system (TOPS). This system is framed as a topological saliency prior (topo-prior)-guided two-level local cue processing (i.e., pixel- and regional-level cues) with a multi-scale strategy, which includes three main modules: (1) a basic computational model of the topological perception theory for extracting topological features from images, (2) a topo-prior calculation method based on the topological features, and (3) a global–local saliency combination framework guided by the topo-prior. Extensive experiments on widely used salient object detection (SOD) datasets demonstrate that our system outperforms the unsupervised state-of-the-art algorithms. In addition, the topo-prior proposed in this work can be used to boost supervised methods including the deep-learning-based ones for fixation prediction and SOD tasks.