A novel SRGAN framework for mural restoration: integrating multi-scale features and attention mechanisms

Intangible cultural heritage, especially Xinjiang murals, eloquently testifies to the vast historical expanse of China, capturing the unique spiritual core and the deep cultural importance that defines Chinese civilization. Unfortunately, the unceasing progress of time has caused varying levels of damage to these irreplaceable treasures. The current techniques for mural restoration face issues such as unclear textures and a lack of specific, targeted solutions. This study introduces an advanced mural restoration model that utilizes generative adversarial networks, known as SR-GAN. The model includes a multi-stage generator that is divided into tasks for global and detailed restoration, allowing for a more thorough examination of both the minute details and the broader aspects of the murals. By employing spatial transformation operations and a separable attention mechanism, along with introducing a new channel and spatial attention mechanism, the network gains cutting-edge capability to identify the complex spatial connections between features. The experiments conducted in this study utilize a bespoke Xinjiang Beiting mural image dataset for training and testing. The results demonstrate that our proposed algorithm achieves an average enhancement of 1.84 dB in peak signal-to-noise ratio (PSNR) and an average improvement of 3% in structural similarity index (SSIM). The results of this research can be effectively applied to the conservation of cultural heritage such as murals.