A plug-and-play intra-class variance suppression framework for industrial anomaly detection

Anomaly detection methods leveraging unsupervised learning are expected to find broad application across diverse sectors, especially in inspecting defects of industrial products. This potential is largely due to their resilience against the unpredictability of anomaly types and the imbalances of learning data across classes. Central to these methods is the premise that feature extractors or image reconstructors, when trained solely on normal data, are incapable of fully replicating the features or inputs of anomalous data. As a result, anomalies could be detected by thresholding the deviations in the extracted features or the reconstructed outputs. However, finding an optimal threshold that effectively separates anomalous from normal data remains a substantial challenge in real-world scenarios. The inherent variability within normal data itself is a significant factor contributing to this challenge. In this study, we introduce a simple yet powerful intra-class variance suppression framework that enables anomaly detection models to suppress intra-class variability by learning compact representations of normal data. We evaluate the proposed framework on three established unsupervised anomaly detection paradigms, namely generative adversarial learning, knowledge distillation, and reverse distillation. Experiments are conducted on multiple benchmark datasets, including handwritten digit images, natural object images, industrial anomaly detection benchmarks, and two additional real-world industrial datasets. The results demonstrate that the proposed framework consistently improves anomaly detection and localization performance, particularly in practical industrial quality inspection scenarios.