Dual-domain feature extraction and reference image guidance for large-area digital mural restoration

Digital mural restoration has gained significant attention due to its importance for cultural heritage preservation and the rapid advancements in deep generative models. However, existing methods often struggle to restore large, heavily damaged regions, leading to suboptimal content generation. We argue that these limitations stem from ineffective feature extraction and the lack of reference information in severely damaged murals, resulting in insufficient understanding of the known content. To address these challenges, we propose a novel restoration approach for large-area mural damage, leveraging dual-domain feature extraction and reference image guidance. First, we introduce a novel frequency-domain representation to extract high-frequency textures, enhancing the global receptive field to complement existing local feature extraction methods. Then, we incorporate reference images as prompts to guide the restoration process. A coarse-to-fine matching and fusion strategy is proposed to effectively align the input damaged mural with the reference image, facilitating the extraction of candidate features. Finally, the restored mural is generated with improved structural integrity, rich textures, and harmonious color coordination. Additionally, we have constructed a mural image dataset with reference image pairs based on their relevance. Extensive experimental results across multiple datasets demonstrate the superiority of our method in large-area mural restoration.