Multiscale voxel feature fusion network for large scale noisy point cloud completion in cultural heritage restoration

The high-fidelity digitization of large-scale cultural heritage via point clouds is often hindered by occlusions and environmental noise during scanning, compromising morphological analysis and restoration. To address challenges such as noise interference and inadequate detail reconstruction, we propose a novel three-stage framework. It incorporates a Multistage Filtering Module to cleanse raw point clouds by balancing noise suppression with geometric preservation. A Multiscale Voxel Feature Fusion Framework hierarchically extracts and fuses features at varying voxel granularities, enhancing recovery of both global structures and local details. Additionally, a Curvature-guided Feature Enhancement Module sharpens reconstruction in high-curvature areas during skeleton point prediction. Experiments show our method outperforms existing approaches by 16% on the ShapeNet-55 and 12% on a real-world cultural heritage roof dataset. The application to Tamaki-jinja Shrine data effectively completed the missing roof regionsc and transparent visualization results confirmed improved perceptual clarity and structural visibility, thereby validating its utility for digital preservation.