Emphasizing Boundary-Positioning and Leveraging Multi-scale Feature Fusion for Camouflaged Object Detection

Songlin Li, Xiuhong Li, Zhe Li, Boyuan Li, Chenyu Zhou, Fan Chen, Tianchi Qiu, Zeyu Li

Abstract: Camouflaged object detection (COD) aims to identify objects that blend in with their surroundings and have numerous practical applications. However, COD is a challenging task due to the high similarity between camouflaged objects and their surroundings. To address the problem of identifying camouflaged objects, we investigated how humans observe such objects. We found that humans typically first scan the entire image to obtain an approximate location of the target object. They then observe the differences between the boundary of the target object and its surrounding environment to refine their perception of the object. This continuous refinement process helps humans eventually identify the camouflaged object. Based on this observation, we propose a novel COD method that emphasizes boundary positioning and leverages multi-scale feature fusion. Our model includes two important modules: the Enhanced Feature Module (EFM) and the Boundary and Positioning joint-guided Feature Fusion Module (BPFM). The EFM provides multi-scale information and obtains aggregated feature representations, resulting in more robust feature representations for the initial positioning of the camouflaged object. In BPFM, we mimic human observation of camouflaged objects by injecting boundary and positioning information into each level of the backbone features, working together to refine the target object in blurred regions progressively. We validated the effectiveness of our model on three benchmark datasets (COD10K, CAMO, CHAMELEON), and the results showed that our proposed method significantly outperforms existing COD models.