MG-AD: Mask-Guided Student-Teacher Training for Real-Time Anomaly Detection

Advances in the latest technologies such as Internet-of-things (IoT), 5G, artificial intelligence (AI) and edge computing have brought the manufacturing industry to a new level of automation. Among such automated tasks, anomaly detection has emerged significantly due to its wide applicability in various tasks such as failure inspection and quality control. Anomaly detection must be both real-time and accurate, yet it is challenged by the scarcity of anomalous data and model sizes that exceed edge-device budgets. In this work, we propose MG-AD, a mask-guided student-teacher framework that introduces the usage of not only synthesized images but ground truth masks from such images, addressing limitation of standard synthetic anomaly generation methods and anomalous data shortage. To address real-time response and deployment to devices with limited capabilities, MG -AD is trained through the knowledge distillation method and features Cramming -Net, a lightweight yet powerful feature extractor, as the student network. With the stated contributions, MG-AD delivers 1.49 ms latency on MVTec AD, with 98.3% AUROC and 89.1% AUPRO for detection and localization, respectively.