AMF-Net: adaptive multiscale fusion network for high-precision camouflaged object detection

The primary objective of Camouflaged Object Detection is the identification and segmentation of entities that are visually integrated into their environment via consistent chromatic and textural patterns. We propose AMF-Net, featuring an Adaptive Multi-scale Fusion Module (AMFM) to bridge the semantic gap. By employing a dual-branch depthwise separable convolution strategy, AMFM decouples texture extraction from global context. Enhanced by dynamic weighting and residual gating, the model effectively suppresses noise and sharpens boundaries. Experiments on COD10K, CAMO, and NC4K show AMF-Net significantly outperforms SOTA methods, achieving Sα = 0.884 and MAE = 0.020 on COD10K with high computational efficiency.