Structure-Guided Texture Enhancement and Multi-Scale Feature Fusion for Progressive Mural Inpainting

Progressive mural inpainting requires structural guidance to reconstruct missing textures while maintaining color consistency. Existing methods only introduce line constraints at the input stage, resulting in distorted structures and color mismatch for severely damaged murals. This paper proposes SGMA-MuralNet, a two-stage network integrating structure-conditioned deep feature modulation and multi-scale feature aggregation (MFA). A full-line encoder applies affine modulation on channel-local normalized features in middle encoder layers to retain fine mural structures. The MFA module uses parallel dilated convolutions and channel recalibration to capture multi-range context. A color correction network (CCN) predicts signed residual for color alignment, and mask compositing keeps intact original pixels. Evaluated on DhMurals1714 with 10–60% irregular damage masks, our model reaches 28.65 dB peak signal-to-noise ratio (PSNR), 0.926 structural similarity index measure (SSIM) and 0.034 learned perceptual image patch similarity (LPIPS), outperforming MuralNet with +1.37 dB PSNR and 30.6% LPIPS reduction. Visual results verify its capability to restore facial contours and flowing ribbons with coherent color transitions. The contribution is an incremental architectural integration of established structural conditioning and multi-scale aggregation mechanisms within MuralNet’s progressive structure reconstruction network (SRN)–CCN framework. The reported results support its utility for benchmark mural inpainting conditioned on supplied registered drawings; they do not establish the historical authenticity of generated details.