Boundary-aware High-resolution Network with region enhancement for salient object detection

Abstract Existing deep learning based salient object detection methods have achieved gratifying progress, however they still suffer from the coarse boundaries and incompleteness of salient objects. To alleviate these issues, this paper presents a novel method that supplements the low-resolution high-level semantic information to high-resolution low-level information with boundary optimization and region enhancement. Based on the parallel architecture, we design a Boundary-aware High-resolution Network (BHNet). First, BHNet maintains high resolution to extract features of the image at the first pathway, and the resolution of the other four pathways are lower, which are used to provide more semantic information for the first pathway. Second, in order to better integrate multi-level semantic information into the first pathway and improve the ability of the model to perceive salient objects, several Multi-path Channel Weight Modules (MCWMs) and Region Enhancement Modules (REMs) are further designed for corresponding blocks. Finally, we also propose a boundary loss function to guide the network to learn more detailed boundary information, which leads to accurate predictions with clear boundaries. Exhaustive evaluations on 6 popular datasets illustrate that the proposed method outperforms the state-of-the-art approaches due to its superior performance, nice generalization and powerful learning ability.