Reference-Guided Inpainting of Dunhuang Murals via Structure-Aware Construction and Feature Filtering

To address the dual challenges of limited reference availability and inaccurate feature matching in Dunhuang mural inpainting, this paper proposes a reference-guided inpainting framework based on structure-aware construction and feature filtering, named DM-SACF. First, we introduce a Structure-aware Reference Construction (SRC) module, which employs the SIFT-Ref method to match reference images that are most consistent with the target damaged mural in local structure. In this way, the model can construct reliable reference information and alleviate the lack of prior guidance. Second, we incorporate a Reference Feature Filtering (RFF) mechanism into the inpainting network. By applying a Top-k sparse selection strategy, the proposed strategy suppresses irrelevant reference information and improves the matching accuracy between reference features and damaged regions. Extensive comparative and ablation experiments on a constructed Dunhuang mural reference dataset demonstrate that the proposed method consistently outperforms existing mainstream inpainting approaches in both visual quality and quantitative evaluation.