Mural Restoration via the Fusion of Edge-Guided and Multi-Scale Spatial Features

To address the issues of low contrast and blurred edges in Dunhuang murals, which often lead to artifacts and edge-detail distortions in restored areas, this study proposes a mural restoration algorithm via the fusion of edge-guided and multi-scale spatial features (EGMF). First, the mural image is fed into an encoder to extract low-level features. During the encoding stage, an Edge-Gaussian Fusion Block (EGFB) is introduced to enhance edge details. This block leverages the rotation-invariant Scharr filter to preserve high-frequency edge details and employs Gaussian modeling to refine low-confidence features, thereby highlighting salient structures and suppressing background noise. Moreover, during the decoding phase, a Hybrid Pyramid Fusion Mamba Block (HPFMB) is introduced. This decoder applies Dense Spatial Pyramid Pooling (DSPP) to aggregate semantic information across multiple scales, and a Pyramid Fusion Mamba (PFM) Module further reduces redundant semantics, thereby enhancing multi-scale feature expressiveness and enabling precise detail reconstruction. Finally, the extracted spatial features are delivered to a Spatially Enhanced Mamba Module, which captures long-range dependencies within the state space and performs pixel-level modeling of the damaged murals. Restoration experiments on the publicly available Dunhuang mural dataset demonstrate that the proposed algorithm significantly enhances edge details and enriches multi-scale semantic representations. Specifically, it increases PSNR by 0.04–0.67%, improves SSIM by 0.71–0.84%, reduces L1 error by 10.47–20.95%, and lowers the LPIPS metric by 1.18–14.45%.