Camouflaged object detection based on multi-scale fusion network with fuzzy region enlargement and shrinking

This paper proposes a novel multi-scale fusion network based on fuzzy region enlargement and shrinking (SVFNet) for camouflaged object detection (COD) tasks. Camouflaged objects are highly similar to their backgrounds in terms of texture, color, and shape, making them difficult to detect. To improve detection performance, SVFNet mimics human visual habits by enlarging and shrinking the original image to extract multi-scale features, thereby enhancing detection accuracy. SVFNet first scales the image down to 0.5x and enlarges it to 1.5x, generating multi-scale images, and then extracts features using a shared triple-feature encoder. The DS-ASPP module is used to fuse multi-scale features, while the MAF module aggregates key features through an attention mechanism, and the HFI module strengthens information interaction between channels. SVFNet also introduces an Uncertainty-Aware Loss (UAL) to enhance detection ability in fuzzy regions. Experiments show that SVFNet outperforms existing methods on the CAMO, COD10K, and NC4K datasets, particularly in low-contrast and complex backgrounds, reducing false positives and false negatives. Ablation experiments validate the effectiveness of each module. Although the model could be improved in terms of middle-layer feature handling and training efficiency, SVFNet provides an innovative and efficient solution for camouflaged object detection.