Signal, Image and Video Processing Обработка сигналов
SCMA-Net: structure-consistent mask-aware network for high-fidelity mural inpainting
Аннотация
Mural image inpainting requires structural continuity and fine-grained texture reconstruction. We propose the Structure-Consistent Mask-Aware Network (SCMA-Net), a hierarchical encoder–decoder framework combining Damage-Aware Feature Downsampling (DAFD) and Structure–Texture Complementary Blocks (STCBs). DAFD preserves feature–mask correspondence through joint rearrangement and interleaving, followed by spatially varying, channel-dependent gating. Within STCBs, Multi-Branch Dependency Modeling (MBDM) integrates channel-dependency attention and dilated spatial context. Shared spatial modulation and selective channel modulation provide asymmetric feature fusion. Experiments on simulated mural damage show improved reconstruction fidelity, structural similarity, and perceptual quality over representative inpainting methods. Qualitative comparisons and a blind user study on authentically damaged murals further support the visual coherence of the reconstructions.