Exploring Generalization Boundaries of Unsupervised Industrial Anomaly Detection Models through Attribute Perturbation

Industrial anomaly detection (IAD) plays a crucial role in large-scale industrial manufacturing. Recently, numerous unsupervised algorithms have been proposed and achieved remarkable performance on benchmark datasets. Given the high homogeneity of samples during training and testing, it appears that the current datasets have been exhaustively addressed. However, it remains uncertain whether these state-of-the-art (SOTA) methods can perform well in more diverse real-world industrial scenarios, or merely overfit to the existing datasets. To address this issue, we propose an attribute perturbing framework utilizing foundation models. Based on it, we quantitatively analyze the impact of attribute perturbation on the anomaly detection system. To the best of our knowledge, we are the first to evaluate the generalization ability of current IAD methods under a shift in testing conditions Systematically. This work helps us have a deeper understanding of the limitations in IAD methods, which also offers valuable insights for future dataset building.