Attention-induced semantic and boundary interaction network for camouflaged object detection

Abstract

The high intrinsic similarity between camouflaged objects and the background makes camouflaged object detection (COD) more challenging than traditional salient object detection (SOD). Existing deep-learning methods often fall into the following shortcomings: (1) When dealing with high-level features, it is difficult to fully and accurately extract semantic information, which is crucial for locating camouflaged targets. (2) The camouflaged targets identified by existing methods in complex scenes have rough boundaries and incomplete spatial information. (3) Existing methods cannot effectively exert the uniqueness and complementarity of multiple features when integrating multiple features. To this end, we propose an attention-induced semantic and boundary interaction network for accurately identifying camouflaged objects. Specifically, we propose a contrastive positioning module (CPM), which adopts a contrastive learning way to separate the camouflaged object from the background, thereby obtaining the exact location of the camouflaged target and retaining more semantic information. Then, we design a boundary exploration module (BEM), which can reduce background noise interference, focus on the structural details of camouflaged targets, and explore rich fine-grained spatial information. Finally, we propose an attention-induced interaction module (AIM) to fuse multivariate information (i.e., semantic location information, boundary information, and the side output information of the backbone), and multivariate information is complementarily fused under attention guidance to generate more powerful camouflaged target features. Extensive experiments show that the proposed method on the different backbones (i.e., ResNet-50, Res2Net-50, PVTv2-B2, EfficientNet-B1) produces state-of-the-art results on several camouflaged benchmarks. The code is available https://github.com/zhangqiao970914/ASBI.