Abstract:

Mural digital inpainting aims to employ information technology to fill the missing areas of damaged murals, thereby restoring their visual integrity and original artistic appearance.To address the limitations of existing deep learning methods for ancient mural image inpainting, which often solely rely on internal priors and lack external feature guidance, resulting in semantic inconsistency and blurred details, this paper proposes a reference image guided and style enhanced ancient mural image inpainting method. First a backbone network for mural image inpainting is constructed, along with a mural feature encoding module based on adaptive cross-scale convolution, which extracts mural content features across multiple scales to enhance the model’s ability in restoring fine-grained details. Next, a style feature encoding module is designed to learn and extract style features at different scales from reference mural images. Then, a feature alignment and fusion module is introduced to align and integrate the content features extracted by the encoder with the multi-scale style features from the reference image, enabling external style-guided information to participate in the inpainting process. Following this, a style-aware enhancement module is constructed to further refine the fused style features. Meanwhile, a dynamic feature guidance layer is designed within the decoder of the inpainting network to guide the restoration process, improving the semantic consistency of the final output mural. Finally, restoration experiments on the Dunhuang mural dataset were conducted, and quantitative analysis was performed using PSNR and SSIM as objective evaluation metrics. The results demonstrate that the proposed method can effectively restore damaged murals, achieving superior performance in both objective and subjective evaluations compared with existing methods.

Key words:

mural image inpainting, reference image guidance, style enhancement, feature alignment and fusion, guided inpainting