Deep Learning-Based Anomaly Detection in Industrial Images: Evaluation and Comparison of Modern Methods

Unsupervised anomaly detection (AD) plays a key role in industrial quality control, enabling automated early detection of defective products without the need for collecting extensive amounts of labeled anomalous training data. However, practitioners and researchers are challenged by choosing suitable AD models in realistic use cases. In this work, we therefore evaluate three modern unsupervised AD models, namely EfficientAD , PatchCore , and FastFlow , on (1) the well-established MVTec AD dataset and (2) the new real-world Rubber Mats dataset, which contain subtle and diverse defects. In a design-science-oriented experimental study, the AD models selected are benchmarked. Further, the impact of higher image resolutions is evaluated, and generalization in multi-class training across categories of instances is examined. All models achieve high overall detection performance but differ in strengths: EfficientAD-M balances discriminative performance and efficiency best, PatchCore benefits from higher resolutions but at the cost of efficiency, and FastFlow generalizes well with stable detection performance. Our findings emphasize the importance of selecting models based on application-specific needs. The results offer practical insights for deploying visual AD systems in real production environments.