Mural image restoration with spatial geometric perception and progressive context refinement

Abstract

Ancient murals, as invaluable cultural heritage, have long been a focal point and significant challenge in the field of cultural heritage preservation. Traditional restoration methods typically address texture and structural features separately, leading to inconsistencies between local details and the overall structure. This approach is insufficient to meet the complex demands of texture and structural restoration for ancient murals. To address this issue, this paper proposes a collaborative encoder–decoder architecture (MIR-SGPR) that achieves simultaneous restoration of texture and structural features in ancient mural images. The generator extracts shallow texture features and deep structural features through the encoder and, in conjunction with the Spatial Geometric Awareness (SGA) module, achieves precise modeling of the spatial location and directional information of damaged areas. To resolve the imbalance between local details and global semantics, this paper introduces the Progressive Contextual Refinement (PCR) network, which progressively optimizes multi-scale features and effectively integrates texture and structural information, thereby enhancing the collaborative modeling capability of local details and global structure. Furthermore, this paper proposes the Mask Reverse-Focus Mechanism (MRF), which leverages mask information to eliminate feature interference from undamaged areas, significantly improving the efficiency and accuracy of restoration. Ultimately, the generated images are optimized through both the global and local discriminators. Experimental results demonstrate that this method significantly outperforms existing state-of-the-art approaches across multiple evaluation metrics. The generated restored images exhibit superior visual consistency, detail authenticity, and overall structural recovery, providing an efficient and reliable solution for the digital preservation of ancient murals.