MuralRescue: Advancing Blind Mural Restoration via SAM-Adapter Enhanced Damage Segmentation and Integrated Restoration Techniques

In this paper, we introduce an innovative method for blind mural restoration, named “MuralRescue,” which demonstrates a systematic approach to progressively restore and enhance the quality of Dunhuang mural images by integrating damaged area segmentation, inpainting processing, and super-resolution techniques. In the process of mural damage segmentation, we employ the SAM-Adapter to optimize the “Segment Anything” model and to enhance the performance of mural damage segmentation. Specifically, we use an adapter module containing two layers of MLP to fine-tune the “Segment Anything” model, thereby increasing the accuracy of segmenting mural cracks. Through extensive experiments, we have proven the adapter’s effectiveness in detecting small targets and fine-grained mural cracks. Additionally, by combining detected cracks with image restoration, we have significantly improved the superiority of blind image restoration tasks without reference.