Abstract:To address the issues of unreasonable structural inpainting and insufficiently refined image texture features in the inpainting of damaged Dunhuang mural images, this paper proposes a cascaded iterative fine-grained mural image inpainting model. This model utilizes a multi-layer residual-net (MR-Net) combined with spatial and channel attention mechanisms,and employs an iterative layernorm residual module (LRM) together with standard convolution for initial inpainting, progressively extracting both global and local feature information of the mural images. Subsequently, the dsconv layernorm teansposed-attention (DSLT-A) and the gated feature aggregation propagation module (FAPM) are combined to reconstruct the global and local texture features. A partial convolution multi-scale gating module is used to enhance missing original image feature information during the inpainting process, refining the texture information of the mural images. Experiments were conducted using artificially irregular mask inapinting, simulated mold spot mask restoration, and real damaged mural inpainting. The experiment results show that, under three objective evaluation metrics, compared with the best-performing data among four comparative models, the model proposed in this paper improves the peak signal-to-noise ratio of the inpainted images by 1.178 dB, structural similarity index by 0.010 2, and reduces mean squared loss by 0.000 9, all outperforming the other comparative models. The proposed model is capable of inpainting higher-quality mural images, providing technical support and reference for improving the inpainting level of Dunhuang mural images and promoting future research in digital inpainting of Dunhuang murals.