Design of Camouflaged Object Detection Algorithm Based on Skeleton Guidance

The main challenge in the camouflaged object detection (COD) task lies in the fact that the targets and backgrounds exhibit a high degree of texture and color similarity. Inspired by human's habit of observing objects, this paper designs a three-level structure network combining skeleton information, including three stages: observation and search, fusion and amplification, judgment and segmentation. In the observation and search stage, an attention pyramid vision Transformer network is used to locate objects. In the fusion and magnification stage, a Fusion Expansion Module (FEM) is utilized to fuse and amplify features from adjacent layers in the feature map. In the judgment and segmentation stage, the Alternating Attention Module (AAM) achieves precise separation by alternately focusing on highly similar foreground and background, completes the segmentation of camouflaged targets. The skeleton information helps to obtain the geometric structure and extended attributes of objects, improving the accuracy of camouflaged object segmentation. Experiments on benchmark datasets show that the model performs well not only on Camouflage Object Detection (COD) tasks but also on downstream COD tasks, and can be applied to camouflaged object detection and medical polyp detection.