Inpainting of Dunhuang murals based on progressive reference guidance and sparse feature matching

Reference-guided inpainting for Dunhuang murals is limited by unreliable reference selection, excessive dependence on ideal references during training, and interference from irrelevant reference features. To address these issues, a Dunhuang mural inpainting method based on progressive reference guidance and sparse feature matching (PRGSM) is proposed. Firstly, a multi-level similarity-based reference construction method (MSR) retrieves references from structural, perceptual, and semantic perspectives, providing high-confidence priors. Secondly, a progressive reference-guided training strategy (PRT) uses varied-quality references across training stages to reduce early over-reliance on ideal references, enhancing robustness and inpainting quality. Finally, a reference Transformer module based on the joint sparse reference matching method (JSR) combines a top- k sparse selection mechanism with a temperature scaling modulator to suppress irrelevant reference responses and improve feature matching. Experiments on our DHRef dataset demonstrate that PRGSM outperforms mainstream methods in both visual quality and objective metrics, particularly for complex patterns and large missing regions.