A two-stage anomaly detection framework: Towards low omission rate in industrial vision applications

Abstract

In industrial manufacturing, there are many types of defective samples that are difficult to obtain. Practical industrial vision anomaly detection has proven to be a challenging task because techniques use only normal (non-defective) samples to train a model to detect anomalies. Currently, some reasonably effective models do not perform very well once differences between samples are large, and they ignore the fact that the cost of missing a defect is much higher than the cost of misidentifying a normal sample. To that end, in this paper, we propose a two-stage framework to construct an anomaly detector. We first train a classification network and then build a one-class classifier on learned representations using another pre-trained network. This paper innovatively proposes using the theoretical quantile as the discriminant threshold. We conduct experiments on the Nut and Motor Brush Holder datasets from real industrial production lines. The results show that our method greatly reduces missed detection of anomalous samples, achieving state-of-the-art AUROC scores of 99.3 % and 96.2 %. We also conduct experiments on the publicly available dataset Rd-MVTec AD, showing that our model has good generalizability and fast testing speed while maintaining high AUROC scores. Our model gives excellent results for nonaligned and defective data with diverse anomalous patterns, and it is easy to optimize. Therefore, not only does our technique handle industrial cold starts well, but it also meets the requirement of online updating, which indicates that our solution is highly suitable for industrial manufacturing scenarios.