Adaptive Perception and Filtering Decoding for Edge Detection Diffusion Models

Mural sketch extraction is an important component in the digital preservation of Dunhuang murals. However, due to physical deterioration over time, the murals exhibit complex background noise and blurred, fragmented lines. It leads to frequent misdetections and omissions in traditional methods. To address this issue, we propose a diffusion-based mural line extraction method that is not limited by image resolution. The method accurately distinguishes true structural edges from background interference, enabling near single-pixel-level high-quality line art generation.Experimental results show that the method achieves the best performance on both the self-built Dunhuang mural dataset and the public BIPED dataset. Also, this method demonstrates that it is effective in real mural scenarios as well as strong generalization ability. The generated fine-grained line art provides high-fidelity structural support for subsequent mural digital reconstruction.