Structure prioritized mural inpainting with edge guided refinement and CNN Transformer fusion

Murals suffer from degradation disrupting structural and textural coherence. Existing inpainting methods entangle structural reconstruction with detail synthesis, causing boundary drift in large missing regions. We propose a structure-prioritized progressive framework leveraging explicit edge priors for mural inpainting. A Structure Reconstruction Network establishes a coarse scaffold, followed by an Edge-guided Refinement Network for boundary-aware correction. Specifically, an Edge-Guided Cross Attention module explicitly injects structural priors into multi-scale interactions, while a CNN-Transformer Fusion Block models local textures and long-range dependencies. On DhMurals1714, our method achieves 24.97 dB PSNR, 0.8558 SSIM, and 0.0738 LPIPS under a severe 40% mask ratio, with 16 ms inference per image. Experiments on authentic murals and simulated specific pathologies confirm its effectiveness, providing a highly controllable, non-invasive solution for digital cultural heritage preservation.