Using High-Quality Feature for Weakly-Supervised Camouflaged Object Detection

Camouflaged objects appear almost identical to the background, which leads to low-level features carrying a large amount of noise, and redundant information being included in high-level features. Therefore, we propose the Detailed and Semantic Feature Extraction Network (DSNet) to recognize camouflaged objects from the background. Firstly, we use the Top-k Sparse Attention module (TKSA) to avoid interference from noise in low-level features. Secondly, we introduce the Spatial and Channel reconstruction Convolution module (SCConv) to remove redundant information in high-level features, at the same time, we propose the Asymmetric Logical Semantic Relation module (ALSR) to capture more semantic information in high-level features. Finally, extensive experiments show that our method achieves significant improvements compared to other weakly-supervised camouflaged object detection methods on three datasets.