Evaluation of deep unsupervised anomaly detection methods with a data-centric approach for on-line inspection

Alexander Zeiser, Bekir Özcan, Bas van Stein, Thomas Bäck

Abstract: Highlights•Comparison of anomaly detection methods for discovery of unlabelled defects in AM.•Utilisation of domain randomization to abstract from real to synthetic image data.•Concept of synthetic data in-the-loop to improve experimental efficiency.•Implementation of a proof of concept for online anomaly detection based on WGAN.•Detection of unknown anomalies with >99% accuracy evaluated on synthetic data.