Abstract
This paper addresses the problems of fading, loss, and texture degradation in the Zuojiang Huashan rock art caused by long term weathering and human disturbances, and constructs a Huashan rock art dataset for image restoration while proposing a prior guided restoration framework based on Stable Diffusion. The proposed method introduces a dual branch prior extraction module to separately model structural contours and texture material information during the diffusion denoising process, and employs a gate attention fusion module to achieve adaptive prior injection and weighted fusion, thereby improving structural continuity and material consistency in missing regions. At the data level, a total of 528 candidate images are collected, among which 177 valid samples are retained after unified screening and cleaning, and training and evaluation pairs are constructed using a unified 512×512 preprocessing scheme together with a mask generation strategy. The experiments are evaluated using MSE, MAE, PSNR, and SSIM, and the results show that the proposed method achieves an MSE of 0.021, an MAE of 0.093, a PSNR of 26.73, and an SSIM of 0.892 on the test set, outperforming multiple representative baseline methods in terms of overall performance. Ablation studies further verify the effectiveness of the dual branch prior extraction and gate attention fusion modules, with the complete model consistently improving all four metrics compared with the Stable Diffusion baseline. In addition, paired significance tests and focused region visualizations demonstrate that the performance gains are statistically meaningful and reveal that the model primarily focuses on missing boundaries, key symbolic contours, and texture transition regions, providing support for interpretable restoration in cultural heritage scenarios.