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

Area of Science:

• Industrial Engineering
• Computer Vision
• Machine Learning

Background:

• Unsupervised defect detection is crucial for industries to avoid complex fault sample collection.
• Existing methods struggle with distinguishing normal from abnormal conditions, leading to high false-positive rates.
• False alarms increase workload and hinder the adoption of unsupervised anomaly detection.

Purpose of the Study:

• To develop a novel method for reducing false positives in unsupervised industrial defect detection.
• To enhance the reliability and practical applicability of unsupervised anomaly detection models.

Main Methods:

• Introduced the False Alarm Identification (FAI) method, utilizing anomaly-free images.
• Employed a multi-layer perceptron to capture semantic information of potential false alarms.
• FAI functions as a post-processing module, filtering predictions from baseline detection algorithms like normalizing flows.

Main Results:

• The FAI method effectively identifies and filters out false alarms generated by unsupervised defect detection algorithms.
• Demonstrated significant reduction in spurious alerts across extensive industrial applications.
• Validated effectiveness when integrated with state-of-the-art normalizing flow algorithms.

Conclusions:

• The FAI method significantly improves the precision of unsupervised defect detection systems.
• By reducing false positives, FAI facilitates wider adoption of anomaly detection in industrial settings.
• This approach offers a practical solution for enhancing the reliability of automated inspection systems.