Damaged mural image inpainting based on reference-guided multi-scale attention mechanism

The restoration of Dunhuang mural images not only protects and preserves Chinese culture but also advances academic research and development in related ffelds. To address the limitations of existing methods in the restoration of damaged murals, such as inadequate feature extraction, unrealistic texture detail restoration, and structural inconsistencies, this paper proposes a novel inpainting method for damaged mural image based on reference-guided multi-scale attention mechanism, referred to as RGMAM. Firstly, we design the reference-based encoder-decoder network to extract texture and structural features from damaged and reference mural images. Then, a multi-scale attention feature alignment module aligns and fuses these features at the pixel level. Subsequently, a feature equalization strategy is applied to the repaired regions to enhance texture and structural information in the fused feature maps. Finally, a pyramid dilated convolution residual block further strengthens contextual representations. Experiments on simulated and real mural datasets demonstrate that RGMAM achieves superior performance over state-of-the-art methods in both quantitative metrics and visual fidelity, producing coherent restorations with clear textures and consistent structures.