An improved mural image restoration method based on diffusion model

Murals, as a vital component of cultural heritage, face severe threats from natural aging and human-induced damage. Given their unique cultural and historical value, restoring damaged murals is not only an artistic endeavor but also a significant act of heritage preservation. Although denoising diffusion probabilistic models (DDPM) have shown remarkable performance on public image datasets, mural restoration remains a challenging task. This paper proposes an improved U-Net-based noise estimation network within the DDPM framework, incorporating a hybrid attention mechanism and an atrous spatial pyramid pooling (ASPP) module, referred to as U-Net-hAM-ASPP. Experimental results demonstrate that the proposed method achieves superior performance in mural image restoration tasks, significantly enhancing subjective visual quality and yielding notable improvements in objective metrics such as PSNR and SSIM.