Attention-Enhanced Feature-Based Point Cloud Completion Network for Precision Parts

When acquiring point cloud data of precision parts using 3D scanning devices, occlusion or equipment limitations often lead to sparse and incomplete data, resulting in the distortion or loss of key geometric features. To address this issue, this study proposes an attention-enhanced feature-based point cloud completion network for precision parts, using precision bearing rings as an example to construct a dedicated completion dataset for training. The proposed network adopts an encoder–decoder architecture. In the encoder stage, a curvature-weighted sampling feature extraction module and spatial attention mechanism are introduced to extract both local and global features from the incomplete point cloud, followed by multilevel feature fusion. The multiscale features extracted by the encoder are then fed into the decoder, which hierarchically and progressively predicts the missing regions of the point cloud. Finally, an adversarial generation module incorporating a biased attention mechanism enhances the sensitivity of the network to geometric structural differences, thereby producing a complete and refined point cloud as the final output. Experimental results show that on the ShapeNet-part dataset, the proposed network achieves average CD, Pred → GT, and GT → Pred errors of 4.663, 2.459, and 2.457, respectively, representing reductions of 10.8%, 4.7%, and 8.1%, respectively, compared with the mainstream PF-Net completion network. On the bearing ring dataset constructed in this study, the average CD, Pred → GT, and GT → Pred errors were 0.497, 1.064, and 0.601, respectively, decreasing by 9.3%, 16.3%, and 16.2%, respectively, relative to PF-Net. Moreover, the proposed network effectively completed the point clouds of various missing parts, demonstrating its robustness across different types of precision parts.