Yongle Palace murals’ image restoration using deep learning optimization algorithms

Abstract

This study uses generative adversarial networks to address the issues of weathering and pigment fading in the murals of Yongle Palace. It also introduces various attention mechanisms for comparison and takes into account structural information, texture information, and directional attention mechanisms for optimization. For Yongle Palace paintings, a deep learning optimization algorithm-based picture restoration technique is developed. The study’s findings show that the restoration quality is satisfactory and that the qualitative analysis approach may obtain an average subjective opinion score of 4 or above in a variety of mural restoration jobs. In the instance of actual mural damage, the study approaches in quantitative analysis correlate to an average absolute error of 0.016, an average signal-to-noise ratio of 26.234 dB, and a structural similarity of 0.842. The above results indicate that research methods can promote the sustainable development of digital protection of cultural heritage.