Research on the restoration and style transfer model based on improved CycleGAN for Dunhuang murals
Cancan Li, Benli Peng · 2025
When used in image restoration and style transfer, most of the current generative adversarial network models are only suitable for style transformation with vivid texture and color. Due to its limited geometric transformation ability and unstable training, the current generative adversarial network models are not effective for the restoration and style transfer of Dunhuang murals with Chinese traditional art style. In order to solve the problem of difficult preservation and restoration of Dunhuang murals and promote traditional Chinese culture, this paper created a dataset with 430 Dunhuang murals, proposed an improved CycleGAN model, and modified the CycleGAN model by removing identity loss and adding Gram matrix and perceptual loss. And it is applied in the color restoration and style transfer of Dunhuang murals. The experimental results show that the improved CycleGAN model can not only restore the missing colors of Dunhuang murals and restore their original appearance, but also generate more Dunhuang style pictures and enrich their expression forms.