Boundary Information Progressive Guidance Network for Salient Object Detection

In recent years, the use of boundary information in saliency detection has been receiving increasing attention. In some cases, existing methods can output saliency maps with clear object boundaries by learning boundary information. However, their boundary prediction structures are generally separated from the prediction branches of the salient regions, and the resulting boundary features may not match the salient objects. We propose a simple saliency detection unit (SDU) to learn more accurate boundary features, and apply multiple such units to construct a boundary information progressive guidance network (BIPGNet). The SDU cascades the salient region and boundary detections, where the boundary features are directly extracted from the salient regions. In the BIPGNet, semantic and boundary features are progressively merged to produce complementary features. We use the complementary features of each stage in one SDU for detecting the salient objects. In addition, a novel boundary information guidance (BIG) module is designed that focuses on the boundary information in a feature layer. We apply multiple BIG modules to the complementary features at different stages. The quality of output saliency map is improved by modifying the complementary features. Experimental results demonstrate that our method can achieve better performance on five benchmark datasets, consistently surpassing 15 state-of-the-art methods. Our source code is publicly available at https://github.com/CKYiu/BIPG.

Automated summary

This research focuses on the use of boundary information in saliency detection and proposes a new network structure to generate more accurate saliency maps by learning boundary features. Experimental results demonstrate that the proposed method outperforms 15 state-of-the-art methods on benchmark datasets.