Multiprototype representation and controllable anomaly generation for industrial anomaly detection

Abstract

To address the issues of poor generalization with limited samples, lack of directional guidance, and insufficient diversity in anomaly synthesis for unsupervised industrial anomaly detection, this paper proposes a detection method based on multi-prototype representation and controllable anomaly generation. Methodologically, a multi-scale prototype memory bank is constructed to enhance intra-class diversity modeling. Adaptive step-size local synthesis and rectangle-replacement global synthesis are designed to generate texture and structural pseudo-anomalies. A dynamic boundary optimization is further introduced to balance feature compactness and discriminability. Experimental results show that the proposed method achieves 99.2% image-level AUROC and 98.7% pixel-level AUROC on the MVTec AD dataset, with localization accuracy significantly outperforming mainstream methods such as SimpleNet. Moreover, it exhibits strong cross-domain generalization on different datasets.