Restoration of Dunhuang Murals on Large-scale pretraining

Dunhuang murals are a precious cultural heritage and their restoration is of vital importance. Traditional image restoration methods and methods based on generative adversarial networks (GANs) have limitations in the mural restoration task. In this paper, we propose a diffusion model-based method for restoring Dunhuang murals using the RePaint model for image restoration. We first collected and produced 5135 Dunhuang mural image data that had undergone data augmentation methods such as cropping and panning, and then pre-trained the RePaint model on a large public dataset and fine-tunedthe model on the Dunhuang mural data. The experimental results show that the fine-tuned RePaint model has significantly improved the evaluation metrics such as PSNR and SSIM in the Dunhuang mural restoration task compared to other recent excellent image restoration models. This suggests that the RePaint model has superior performance in mural restoration tasks, especially in the areas of texture generation and structure retention. This study provides a novel and effective method for the field of Dunhuang mural restoration, which is expected to provide more support for heritage conservation.