Dunhuang Mural Restoration Based on Generative Adversarial Networks
Chenrui Wang, Yi Fu, Diao Zhou, Ze Shi · 2025
Restoring damaged Dunhuang murals, a vital aspect of preserving cultural heritage, requires methods that effectively address both global structural recovery and fine-grained detail reconstruction. Traditional restoration approaches are labor-intensive and limited in precision, while recent advancements in artificial intelligence, particularly Generative Adversarial Networks (GANs), offer new possibilities for automated and accurate restoration. In this work, we propose a novel GAN-based framework tailored for mural restoration, incorporating the Residual Large Kernel Convolutional (RLKConv) block as a core architectural component. It combines large kernel convolutions with depthwise and pointwise operations, enabling the model to capture global contextual information while preserving fine details. Additionally, the residual connections ensure stable training and efficient feature refinement. The proposed framework demonstrates superior performance in reconstructing missing or degraded mural regions, achieving high visual fidelity and structural consistency.