Exploiting background divergence and foreground compactness for salient object detection

Abstract

In this paper, we propose an efficient and discriminative saliency method that takes advantage of background divergence and foreground compactness. Concretely, a graph is first constructed by introducing the concept of virtual node to effectively enhance the distinction between nodes along object boundaries and the similarity among object regions. A reasonable edge weight is defined by incorporating low-level features as well as deep features extracted from deep networks to measure the relationship between different regions. To remove incorrect outputs, two computational mechanisms are then developed to extract reliable background seeds and compact foreground regions, respectively. The saliency value of a node is calculated by fully considering the relationship between the corresponding node and the virtual background (foreground) node. As a result, two types of saliency maps are obtained and integrated into a uniform map. In order to achieve significant performance improvement consistently, we propose a robust saliency optimization mechanism, which subtly combine suppressed/active (SA) nodes and mid-level structure information based on manifold ranking. Extensive experimental results demonstrate that the proposed algorithm performs favorably against the state-of-art saliency detection methods in terms of different evaluation metrics on several benchmark datasets.