Abstract: As industrial production environments become more intricate, the demand for 3D point cloud industrial anomaly detection is increasing. Although the two-dimensional anomaly detection method based on pre-trained feature extractor is effective, the generalization ability of three-dimensional point cloud pre-training network is limited, which leads to the poor effect of this kind of point cloud anomaly detection method. To bolster the performance of three-dimensional point cloud anomaly detection, Point-ReAD, an innovative method based on point cloud reconstruction, was proposed. Point-ReAD comprises an anomaly simulation module, a point cloud reconstruction network, and an anomaly discrimination module. During training, anomalous point clouds were created from normal point cloud data by the anomaly simulation module, which were then fed into the point cloud reconstruction network, with the normal point clouds serving as self-supervised signals to guide the learning process. The point cloud reconstruction network specifically incorporates the Group Attention Module (GAM), which is meticulously designed to seamlessly integrate intricate structural information from point clouds, thereby enhancing the efficacy of capturing geometric and semantic features in point clouds. In the inference phase, the test point clouds were introduced into the reconstruction network to generate reconstructed point clouds, and anomalies were identified by the anomaly discrimination module by comparing the pre- and post-reconstruction point clouds. The experimental results showed that Point-ReAD achieves point cloud-level AUROC and point-level AUPRO of 95.49% and 94.66%, respectively, on the MVTec 3D-AD dataset, indicating an overall performance improvement compared to methods like M3DM (Multi-3D-Memory) and 3DRÆM (3D Discriminatively trained Reconstruction Anomaly Embedding Model).

Key words: anomaly detection, self- supervised learning, point cloud, computer vision, reconstruction network

摘要： 面对日趋复杂的工业生产环境，三维点云工业异常检测需求与日俱增。尽管基于预训练网络的二维异常检测方法效果显著，但三维点云预训练网络的泛化能力有限，导致这类点云异常检测方法效果不佳。为提高三维点云异常检测性能，提出一种基于点云重构的异常检测方法Point-ReAD(Point cloud Reconstruction for Anomaly Detection)，它由3个核心模块构成：异常模拟模块、点云重构网络和异常判别模块。在训练阶段，正常点云图经异常模拟模块产生异常点云送入点云重构网络，正常点云作为自监督信号指导重构网络学习。点云重构网络使用分组注意力模块(Group Attention Module，GAM)，该模块设计用于融合点云的复杂结构信息，从而有效地捕捉点云中的几何和语义特征。在推理阶段，测试点云进入重构网络后生成重构点云，异常判别模块比较重构前后的点云从而精确定位异常。实验结果表明，Point-ReAD在MVTec 3D-AD数据集上的点云级AUROC(Area Under the Receiver Operator characteristic Curve)和点级AUPRO (Area Under the Per-Region Overlap)分别达到了95.49%和94.66%，相较于M3DM(Multi-3D-Memory)、3DRÆM(3D Discriminatively trained Reconstruction Anomaly Embedding Model)等方法整体性能有所提升。

关键词: 异常检测, 自监督学习, 点云, 计算机视觉, 重构网络