CADDPM: Contour-Aware Denoising Diffusion Probabilistic Model for Mural Inpainting

Abstract Murals are vital cultural heritage, but aging, environmental erosion, and historical damage cause structural fragmentation and color degradation, while inpainting faces structural discontinuity and stylistic inconsistency. To address these, we propose a contour-aware denoising diffusion probabilistic model (CADDPM) for non-contact digital mural restoration. Sobel operator extracts contours from masks, and a lightweight convolutional encoder derives contour embeddings injected into denoising time-step embeddings, providing shape and scale priors for missing regions. Atrous spatial pyramid pooling at U-Net bottleneck enhances multi-scale context modeling. A multi-task loss integrates pixel, perceptual, style, and edge-alignment constraints to ensure structural accuracy, semantic consistency, and color-texture harmony. Inference uses a progressive sampling constraint with free generation, soft blending, and hard constraint stages to mitigate boundary artifacts. Experiments on MuralDH show CADDPM outperforms DDPM, RePaint, and LaMa in PSNR, SSIM, and LPIPS. This framework offers a controllable, interpretable path for digital mural restoration, supporting documentation, analysis, and long-term preservation.