Abstract:

With the growing demand of industrial automation， 3D point cloud anomaly detection has played an increasingly important role in product quality control. However， the existing methods often rely on a single feature， leading to information loss and accuracy reduction. To address these issues， an unsupervised point cloud anomaly detection method based on multi-representation fusion was proposed， called MRF （Multi-Representation Fusion）. MRF used multi-angle rotation and various coloring schemes to render point clouds into multi-modal images， and employed pre-trained 2D convolutional neural networks to extract rich semantic features. Simultaneously， pre-trained Point Transformer was adopted to extract 3D structural features. After the above， by fusing 2D image semantic features and 3D structural features， MRF was able to capture point cloud information more comprehensively. In the anomaly detection stage， abnormal point clouds were identified effectively by using a method based on positive sample memory banks and nearest neighbor search. Experimental results on MVTec 3D AD dataset show that MRF achieves a point cloud-level AUROC （Area Under the Receiver Operating Characteristic curve） of 0.972 and a point-level AUPRO （Area Under the Per-Region Overlap） of 0.948， significantly outperforming existing methods. It can be seen that the effectiveness and robustness of MRF makes it a highly promising solution for industrial applications.

Key words: computer vision, point cloud, unsupervised anomaly detection, feature embedding, memory bank

摘要：

随着工业自动化需求的不断增长，三维点云异常检测在产品质量控制中扮演着越来越重要的角色。然而，现有方法通常依赖单一特征，导致信息损失和精度下降。因此，提出一种基于多表征融合的无监督点云异常检测方法MRF（Multi-Representation Fusion）。MRF利用多角度旋转和多种着色方案将点云渲染为多模态图像，并使用预训练的二维卷积神经网络提取丰富的语义特征；同时，还采用预训练的Point Transformer提取三维结构特征。之后，通过融合二维图像语义特征和三维结构特征，MRF能够更全面地捕捉点云信息。在异常检测阶段，MRF使用基于正样本记忆库和近邻搜索的方法，可有效地识别异常点云。在MVTec 3D AD数据集上的实验结果表明，MRF的点云级接受者操作特征曲线下面积（AUROC）为0.972，点级区域重叠度（AUPRO）为0.948，显著优于对比方法。可见，该方法的有效性和鲁棒性使它成为工业应用中极具潜力的解决方案。

关键词: 计算机视觉, 点云, 无监督异常检测, 特征嵌入, 记忆库