Dual Attention and Edge Refinement Network for Camouflaged Object Detection

Camouflaged object detection (COD) aims to identify and segment objects similar to the background. However, existing methods suffer from unsatisfactory camouflaged object identifying or segmenting. In this paper, we propose a novel Dual Attention and Edge Refinement Network (DAERNet) to boost the performance of COD. Specifically, a Dual Attention Mechanism (DAM) is proposed to capture the scale diversities of camouflaged objects, which mainly includes Spotlight Attention Module (SAM) and Modulation Attention Module (MAM). The SAM aims to obtain multi-scale refinement features which contain discriminative semantic information, while the MAM aims to obtain the refined features with specific semantic information. Then, we propose a Boundary Extraction Module (BEM) to obtain edge information. Finally, a Feature Aggregation Module (FAM) is designed to fuse the refined multi-scale features obtained by DAM with the help of edge information to achieve accurate camouflage object prediction. Extensive experiments on four datasets demonstrate that the proposed DAERNet performs comparably against other SOTA methods.