MFMSA-GAN: multifrequency multiscale attention enhanced GAN for Dunhuang mural face inpainting

Dunhuang murals, as a significant part of traditional art, hold profound historical and cultural value. The facial images depicted in these murals reflect the political, cultural, and religious beliefs of ancient societies. However, prolonged exposure to natural degradation and human activities has severely threatened their preservation and transmission. Consequently, the application of digital restoration techniques for mural facial images has become increasingly imperative. Despite advancements, existing methods struggle with challenges such as complex textures, color style consistency, and local detail reconstruction. To address these issues, we propose MFMSA-GAN, a mural face inpainting model based on a multi-frequency multi-scale attention (MFMSA) mechanism. By integrating MFMSA into the generator's decoder, our approach enhances inpainting accuracy and detail recovery, while a spatial mask discriminator improves attention to the restored areas. Additionally, we introduce a joint histogram loss to enforce color consistency and fine-grained detail preservation, thereby improving the perceptual quality of the restored images. Experimental results, including comparative, ablation, and visual analysis, demonstrate the superior performance of MFMSA-GAN in Dunhuang mural face images inpainting.