Structure aware transformer for multi source point cloud registration in cultural heritage digital modeling

High-precision 3D digitization is essential for cultural heritage preservation, restoration, and digital modeling. However, multi-source point clouds acquired from complex heritage environments often suffer from low overlap, incomplete structures, heterogeneous sampling, and geometric inconsistencies, making robust registration challenging. Here, we propose a structure-aware Transformer framework for multi-source point cloud registration. The proposed method integrates local geometric structures and global contextual relationships through adaptive attention, enabling reliable correspondence learning under complex conditions. A structure-aware local geometry encoding module and a reliability-guided attention mechanism are introduced to improve feature discrimination and registration robustness. Experiments on benchmark datasets and real-world cultural heritage scenes demonstrate that the proposed method achieves superior registration accuracy and fusion consistency, providing an effective solution for high-fidelity digital heritage reconstruction.