□ Prof. Sang Hyun Park's research team in the Department of Robotics and Mechatronics Engineering at DGIST (President Kunwoo Lee) has developed a logical anomaly detection technology in collaboration with a team from Stanford University. This technology is expected to significantly improve defect detection performance in smart factories by leveraging AI to accurately identify logical anomalies in industrial images.
□ Logical anomalies are data that violate basic logical constraints such as the number, arrangement, or composition of components within an image. Unlike structural anomalies, which can be relatively easily detected by examining partial images, the detection of logical anomalies requires the ability to distinguish among various components throughout the entire image. Previous AI models have recorded poor AUROC[1] scores not exceeding an average of 90% in logical anomaly detection.
□ To overcome this performance limitation, Prof. Sang Hyun Park's research team has proposed a model that first learns to accurately segment each component and then uses that information to perform anomaly detection.
□ Typically, training a segmentation model requires pixel-level labeling, which creates a significant labor issue. To address this, the research team has proposed a few-shot segmentation technique utilizing a small number of ground truths. The images used for model training were combined in the same manner, meaning each image is different, but the number of components or pixels is similar. The segmentation model was effectively trained by minimizing the objective function using histograms. Consequently, the proposed technique demonstrated superior accuracy compared to existing few-shot segmentation techniques.
□ Furthermore, the research proposes a model that utilizes image segmentation information to simultaneously perform logical and structural anomaly detection. It uses a total of three memory banks to effectively calculate anomaly detection scores through comparison with test images.
□ The research team applied this technology to the MVTec LOCO AD dataset, which is among the most challenging logical anomaly detection datasets currently available. Whereas the existing techniques have each recorded performance below an average of 90% in logical anomaly detection, the proposed technique achieved an average performance of 98%.
□ Prof. Sang Hyun Park of the Department of Robotics and Mechatronics Engineering at DGIST said, "This research has dramatically improved performance in logical anomaly detection." Moreover, the professor anticipates that "this technology could significantly reduce the costs associated with defect detection in smart factories."
□ This research was conducted with the support of the Daegu Gyeongbuk Institute of Science and Technology and the Daegu Digital Innovation Promotion Agency. The researchers’ results were published by AAAI, the premier AI conference, in February 2024.
- Ccorresponding Author E-mail Address : [email protected]
[1] AUROC (the area under the receiver operating characteristic curve) is a tool to evaluate a model’s predictive ability. A performance score closer to 100% indicates better accuracy.
