Unraveling False Positives in Unsupervised Defect Detection Models: A Study on Anomaly-Free Training Datasets

Unsupervised defect detection methods have garnered substantial attention in industrial defect detection owing to their capacity to circumvent complex fault sample collection. However, these models grapple with establishing a robust boundary between normal and abnormal conditions in intricate scenarios, leading to a heightened frequency of false-positive predictions. Spurious alerts exacerbate the work of reconfirmation and impede the widespread adoption of unsupervised anomaly detection models in industrial applications. To this end, we delve into the sole available data source in unsupervised defect detection models, the unsupervised training dataset, to introduce a solution called the False Alarm Identification (FAI) method aimed at learning the distribution of potential false alarms using anomaly-free images. It exploits a multi-layer perceptron to capture the semantic information of potential false alarms from a detector trained on anomaly-free training images at the object level. During the testing phase, the FAI model operates as a post-processing module applied after the baseline detection algorithm. The FAI algorithm determines whether each positive patch predicted by the normalizing flow algorithm is a false alarm by its semantic features. When a positive prediction is identified as a false alarm, the corresponding pixel-wise predictions are set to negative. The effectiveness of the FAI method is demonstrated by two state-of-the-art normalizing flow algorithms on extensive industrial applications.