Dense 3D Reconstruction of Ancient Architectural Heritage by Integrating Cross-View Geometric Prior and Semantic Enhancement

Abstract. Ancient buildings are in urgent need of protection due to multiple diseases such as structural crack expansion and painted fading. Multi-view 3D reconstruction of crowdsource images has been widely used due to its high flexibility and low cost. This study proposes a multimodal global optimization framework for the protection of ancient buildings to address the low accuracy of crowdsource image screening and the problems of paired view cumulative error and detail sparsity of the MASt3R method. First, the EfficientNet is improved based on the CBAM mechanism to improve its screening robustness in low-quality crowdsource images. Secondly, a cross-reference view block is designed to achieve cross-view geometric prior fusion through a multi-view attention mechanism. A semantic-guided matching enhancement strategy is further introduced to segment key areas of the building based on the SAM model to focus on detail reconstruction. Experiments show that this method improves the number of point clouds and surface density by 23.4% and 25.9% respectively, effectively solving the problems of geometric distortion and weak texture details in the reconstruction of typical architectural details such as brackets and plaques in traditional methods.