Locate, enhance and fuse: a progressively optimized network for camouflaged object detection

Camouflaged object detection (COD) aims to solve the challenge of blending camouflaged objects with their surroundings. The intrinsic similarity and blurred boundaries between foreground objects and background environments make it difficult to accurately distinguish between the two. Current deep learning methods usually struggle to accurately recognize objects that are camouflaged with complex and fine structures. To address this issue, we propose a novel, Progressively Optimized Network (PONet) for camouflaged object detection. Specifically, we introduce a Contour Guided Module (CGM) aimed at modeling more explicit contour features to improve COD performance. Additionally, we incorporate a Feature Enhancement Module (FEM) with the goal of integrating more discriminative feature representations to enhance detection accuracy and reliability. Finally, we present a Boundary Guided Feature Fusion Module (BGFFM) to boost object detection capabilities and perform camouflaged object predictions. BGFFM utilizes multi-level feature fusion for contextual semantic mining. Subsequently, we incorporate the edges extracted by the CGM into the fused features to further investigate semantic information related to object boundaries, guiding and reinforcing the COD representation learning. By adopting this approach, we are able to better integrate contextual information, thereby improving the performance and accuracy of our model. We conducted extensive experiments, evaluating our PONet method using four challenging datasets. The results show outstanding performance across four widely used metrics.