Dual-Domain Collaborative Modeling With Boundary-to-Structure Refinement for Accurate Camouflaged Object Detection

Abstract

Camouflaged Object Detection (COD) aims to segment objects that blend seamlessly with their surroundings, posing significant challenges due to ambiguous boundaries and incomplete structural perception under complex backgrounds. Existing methods often rely on single-domain representations or coarse feature fusion strategies, which limit their ability to capture subtle boundary cues and holistic object structures, leading to fragmented predictions and susceptibility to background distractions. To overcome these challenges, we propose DDCM, a Dual-Domain Collaborative Modeling framework targeted at robust COD. Specifically, we introduce the Boundary-Aware Dual-Domain Perception Module (BDPM), which combines spatial-domain boundary perception with frequency-domain enhancement to better capture ambiguous edges and their textures. Additionally, to improve recognition of concealed structures under complex camouflage patterns, a Cross-Domain Fusion Block (CDFB) is developed to fuse complementary spatial–frequency cues. Furthermore, the Prior-Guided Decoder (PGD) refines predictions via multi-scale priors and edge-aware fusion, ensuring structural completeness and accurate boundaries. Extensive experiments on three COD benchmarks demonstrate that DDCM achieves leading performance on most metrics, with superior segmentation completeness, boundary precision, and robustness to complex background interference. Furthermore, its strong performance on polyp segmentation datasets suggests promising generalization capability beyond camouflaged scenarios. These results validate the effectiveness of dual-domain collaboration and boundary-guided structural refinement for challenging camouflaged scene understanding.