IDCNet: iterative dual-channel network for camouflaged object detection

Abstract

Aiming at the problem that current camouflaged object detection in military operations, this paper proposes an iterative dual-channel camouflaged object detection network (IDCNet) for this task. This method employs a dual-channel architecture to explicitly assign the tasks of localization and edge refinement, which consist of a robust localization channel and a global refinement channel. Position and channel attention mechanisms are integrated into each high-level feature in the robust localization channel. The localization prediction map is fused with the localization channel features through global attention to obtain the initial features for the global refinement channel. By incorporating a Mirror Multiplicative Attention mechanism and an attention-guided iterative zooming strategy, IDCNet achieves significant improvements in segmentation accuracy. This method not only demonstrates outstanding performance on military camouflaged object datasets but also exhibits excellent performance on general camouflaged object datasets. The model's Structure-measure ($S_{\alpha }$) achieve 91.2% on the military camouflage object dataset CamouflageData. On the largest publicly disguised object dataset COD10K, the $S_{\alpha }$ reach 83.3%. These results underscore the potential of IDCNet to substantially enhance battlefield situational awareness and operational decision-making, paving the way for more robust camouflage object detection in real-world military applications.