Research on structure-guided restoration technique for large-area defects in Regong mural images

<p indent="0mm">The Regong mural is a significant manifestation of the Regong art, a human intangible cultural heritage, characterized by its rich and vivid colors as well as intricate and dense lines. However, prolonged environmental exposure has caused severe degradation, including peeling, cracking, and mildew. Existing automatic image inpainting methods fail to reconstruct their complex structures and generate plausible textures. Addressing the challenges posed by the complex structure and texture of Regong murals, irregular damaged areas, and a limited image dataset, this paper proposes a structure-guided method for restoring large-area defects in Regong mural images. Firstly, an interactive line art restoration module is designed to leverage expert experience in repairing the complex line art of large damaged areas, providing a more accurate structural prior to subsequent steps. Subsequently, the line art is processed through a structure-guided restoration module to avoid weakening low-level guidance information while more effectively integrating with texture information. This module injects encoded structural information into a multi-scale restoration network layer by layer, and an edge adversarial discriminator is designed to constrain and optimize the restoration structure, making it closer to the real mural. Furthermore, the network incorporates Fourier convolution blocks and gated convolution modules to expand receptive fields for irregular large-area damage recovery, and a color consistency loss is introduced to effectively mitigate color discrepancies. Extensive experiments against seven state-of-the-art methods demonstrate that our approach achieves superior performance on murals with 30%–50% damage ratios. Quantitative evaluations reveal improvements of 37.91% in the structural similarity index (SSIM) and 15.3% in the peak signal-to-noise ratio (PSNR), alongside reductions of 94.67% in the Fréchet inception distance (FID) and 85.96% in the learned perceptual image patch similarity (LPIPS).