Robust Industrial Anomaly Detection via Style Shift Estimation and Cascade Distillation

Knowledge distillation (KD) has demonstrated remarkable performance in industrial anomaly detection (IAD) field, but its immense potential is constrained by the trade-off between generalization of normal sample and robustness of abnormal sample. High generalization often weakens the robustness against anomalies, leading to overlooked defects, while enhancing robustness can cause normality forgetting, where normal samples are misclassified as anomalies. This dilemma stems from the existing methods commonly assuming style invariance across normal and abnormal areas, focusing solely on the abnormal content information while style information is usually neglected and put on the shelf. In this paper, our main novelty lies in leveraging anomaly style shift estimation for enlarging abnormal robustness while maintaining high normal generalization. This is the first attempt to introduce the style information into the knowledge distillation-based industrial anomaly detection methods. By employing the bidirectional fusion of style and content information, our approach achieves more comprehensive and practical anomaly detection and localization results. Thus, our method can provide a more balanced and robust solution for the future industrial anomaly detection task. Experimental results on the MVTec2D, MVTec3D, BTAD, and KolektorSDD2 datasets demonstrate that our method outperforms major state-of-the-art (SOTA) methods in both accuracy and speed, with image-level and pixel-level AUROC scores of 99.6% and 98.6% on MVTec2D, respectively. Furthermore, the efficiency of our method is validated in a real-world plastic crate defect detection system on a logistics production line.