A Dunhuang mural restoration network based on mask guidance and Transformer architecture

Abstract In the process of Dunhuang mural restoration, models often struggle to focus on damaged areas, leading to issues such as color deviation and blurred lines. To address these problems, we propose a Dunhuang mural restoration network based on mask guidance and Transformer architecture, named MGTNet. First, we design a mural focus attention module to dynamically optimize both channel and spatial information in an adaptive manner. Next, we introduce a mural enhancement module that leverages the long-range dependency capturing capability of Transformers to improve restoration quality. Finally, a mask-guided downsampling module is proposed, which fuses the mask image as prior knowledge with downsampled features, enhancing the model's ability to perceive damaged areas. Experiments on a publicly available Dunhuang mural dataset demonstrate that the proposed method outperforms comparison algorithms in terms of objective evaluation metrics such as PSNR and SSIM, validating the effectiveness of the algorithm. This study demonstrates the practical application potential of the algorithm in Dunhuang mural restoration and contributes to the significant advancements of artificial intelligence in the field of digital humanities.