Unpaired mural-baimiao generation via gradient-frequency decoupling and structure-guided synergistic attention

Severe weathering and physical deterioration often cause line fading, pigment flaking, and strong edge interference in mural images, making unpaired baimiao generation prone to structural discontinuities, line adhesion, and semantic distortion. To address these problems, we propose a dual-branch framework integrating gradient-frequency decoupling and structure-guided synergistic attention. The Gradient-Guided Frequency-Domain Decoupling Module (GFDM) enhances faint lines while suppressing flaking-induced artifacts, whereas the Structure-Guided Synergistic Attention (SGSA) module uses multi-scale spatial priors to reduce semantic deviations during cross-level feature fusion. A multidimensional joint loss with dynamic total variation weighting further constrains global structure, local details, and background cleanliness. Experiments on a self-constructed unpaired dataset containing 1563 mural images and 1500 baimiao patches show that the proposed method improves artifact suppression and line continuity, producing cleaner and structurally more complete baimiao images. The method provides computational support for mural-baimiao extraction, digital documentation, and conservation-oriented analysis.