Image inpainting using diffusion models to restore eaves tile patterns in Chinese heritage buildings

Abstract

Wadangs (a type of eaves tile) are integral components of traditional Chinese buildings and often suffer damage over time, resulting in the loss of pattern information. Currently, AI-based image inpainting methods are applied in pattern restoration, but face challenges in capturing fine textures and maintain structural continuity. This paper proposes a coarse-to-fine image inpainting method based on the denoising diffusion probabilistic model (DDPM), specifically optimized for wadang pattern restoration. The method starts with an initial inpainting phase, followed by a fusion module that combines the semantic information of the input image with intermediate outputs to achieve refined inpainting results. Experimental results demonstrated that the proposed method outperformed state-of-the-art methods on various evaluation metrics, including PSNR, SSIM, FID and LPIPS, highlighting its effectiveness in restoring wadang patterns by reconstructing damaged areas while preserving the original semantic integrity of the patterns.