Edge-awareness and feature decoupling enhancement network for camouflaged object detection

Camouflaged object detection (COD) aims to detect objects that exhibit similarities in color, texture, and morphology with their surrounding environment, making it a challenging visual task. Existing COD methods have adopted various feature refinement strategies to improve the detection accuracy of camouflaged objects. However, due to the inherent resemblance between camouflaged objects and the background, foreground features are often easily confused with background features, leading to partial or even complete loss of foreground information. To address this problem, we propose a novel edge-awareness and feature decoupling enhancement network (EFDE-Net) for detecting camouflaged objects, enabling more precise segmentation. Specifically, a high-frequency attention module (HFAM) is designed to address the issues of inaccurate object localization and loss of edge detail information caused by unclear boundaries between objects and backgrounds. Then, the edge features and camouflaged features are fully integrated by using the feature fusion module (FFM). Additionally, we design a feature decoupling enhancement module (FDEM) to independently capture and refine subtle features. By decoupling features into strong and weak features, we mitigate the interference of strong semantic features on weak detailed features, thus refining the predictions to achieve more precise results. We conducted both qualitative and quantitative evaluations of our network on three public challenging datasets, including CAMO, COD10K, and NC4K, show that our EFDE-Net presents competitive performance compared with the state-of-the-art models.