Irregularity-Aware 3-D Anomaly Detection for Product Quality Control

Abstract

Point cloud anomaly detection, aimed at detecting subtle geometric deviations, has emerged as a critical task in quality control. While existing methods predominantly employ prototype-based frameworks that quantify anomalies via normal prototypical distances, we identify a fundamental limitation: such approaches neglect the discriminative potential of anomaly prototypes. To address this limitation, we propose Irregularity-Aware 3D Anomaly Detection (IA3D), featuring a novel Refined Prototypical Score (RPS) that synergistically combines absolute and relative distances to both normal and anomalous prototypes. This dual-perspective metric enables refined anomaly quantification through prototype space calibration. To overcome the scarcity of real anomaly data for RPS computation, we design three synthetic anomaly simulation strategies that generate geometrically diverse irregularities. These synthetic samples facilitate the training of an efficient geometric encoder optimized for discriminative local structure representation. Extensive experiments across benchmark datasets (MVTec 3D-AD, Real3D-AD, Anomaly-ShapeNet) and industrial applications demonstrate IA3D’s state-of-the-art performance. Note to Practitioners—Ensuring high-quality industrial parts often hinges on spotting tiny defects during manufacturing. This challenge, especially in precision industries like automotive or aerospace, inspired our work on detecting irregularities in 3D point cloud data. Our method, IA3D, tackles the issue of subtle geometric flaws that current tools often overlook by comparing data to both normal and defective examples. This dual approach catches defects more reliably. Since real defect samples are difficult to obtain, we developed techniques to create synthetic irregular samples for training, making our solution practical even with limited data. Moreover, our fast, lightweight system fits seamlessly into real-time quality checks. It is ideal for industries relying on 3D scans for inspection. That said, the simulated defects might not cover every real-world case, so some additional calibration could be needed. Future efforts could refine these simulations or blend in actual defect data to boost accuracy further.