With the rapid development of manufacturing, surface quality monitoring of products has become important, and it has become important to ensure product quality and production efficiency. Although 2D image-based anomaly detection technology has been extensively researched and applied, 3D scanning technology allows for the capture of full geometric information and high density 3D point cloud data that is not affected by lighting conditions. However, challenges such as complex surface modeling, diverse anomaly types, difficult data representation of 3D point clouds, limited training samples, and diverse and locally sparse anomaly have still lacked effective 3D visual-based surface anomaly detection methods that meet practical accuracy and speed requirements.
Therefore, Huan Du, Xuanming Cao and Fugee Tsung of Hong Kong University of Science and Technology (Guangzhou) and Hong Kong University of Science and Technology jointly conducted a study titled “3D visual-based anomaly detection in manufacturing.” This research is supported by the National Natural Science Foundation of China, Guangzhou Fundamental and Applied Basic Research Foundation, Guangzhou Hkust (GZ) Joint Funding Program, Guangzhou Industrial Information and Information Main Institutes.
This study systematically reviews surface anomaly detection methodology for manufactured products based on 3D machine vision. Specifically, we divide existing anomaly detection methods into three categories: monitored, unsupervised training-based, and untrained methods. For the monitored method, both classic machine learning-based approaches (using local shape descriptors such as point function histograms and rotational projection statistics) and deep learning-based approaches (including point-wise multi-layer perseption, convolutional neural networks, graph neural networks, long-term memory, and transformers) are analyzed. Unsupervised training-based methods can be further divided into functional embedding (such as memory banks and knowledge distillation) and reconstruction-based technologies (such as automatic encoders and principal component analysis). Untrained methods cover local and global geometry-based (statistical and CAD model-based) methods. This study summarizes the strengths, limitations, and application scenarios of each method, highlights long-standing milestone techniques, and provides a detailed comparison with previous related research.
Furthermore, this study brings together existing public datasets for 3D industrial anomaly detection, including MVTEC-3D AD, REAL3D-AD, ANOMALY-SHAPENET, NEU RSDDS-AUG, concrete surface defect dataset, sewer 3D point cloud dataset, data type, potential research type, and latent tone latent type of anomaly specimen. support. It also covers general evaluation metrics for 3D anomaly detection, including accuracy, recall, false positive rate, balanced accuracy, dice coefficients, dice coefficients, dice coefficients, overlap curves per region, overlap curves per region, area under the receiving operating characteristic curve, and the intersection of how it is calculated and significance.
Finally, the study points to future research directions in the field of 3D visual acuity-based anomaly detection in manufacturing. This includes high-precision 3D anomaly detection (focusing on accurate anomaly boundary detection), large-scale language model anomaly detection (developing 3D foundation models, or transferring 2D pretrain models), and anomaly detection based on multi-mass-level data (anomaly detection based on Decturing-level data) methods.
This paper, “3D visual acuity-based abnormality detection in manufacturing: an investigation,” written by Juan Du, Chengyu Tao, Xuanming Cao, and Fugee Tsung. Full Open Access Paper: https://doi.org/10.1007/S42524-025-4189-9.
