Comparison of point cloud completion methods.

With the continuous advancement of 3D scanning technology, point cloud data of mechanical components has found widespread applications in industrial design, manufacturing, and repair. However, due to limitations in scanning precision and acquisition conditions, point cloud data often exhibit sparsity and missing information. This issue is particularly challenging when dealing with mechanically complex geometric shapes, where the missing portions frequently contain crucial details, posing significant difficulties for data completion. To effectively recover these missing parts while maintaining the accuracy of both global morphology and local details, this paper proposes a Hierarchical Shape-Preserving Completion Network (HSPC-Net). This approach integrates a multi-receptive field Transformer with a cross-modal geometric information fusion strategy, enabling the precise restoration of local details of mechanical components at multiple scales. Additionally, it leverages 2D image information to assist in the completion of 3D point clouds, significantly enhancing completion accuracy and robustness. Experimental results on ShapeNet and mechanical component point cloud datasets demonstrate that HSPC-Net outperforms existing state-of-the-art methods in terms of completion accuracy, structural consistency, and detail recovery.