Image Anomaly Detection via Reverse Informative Distillation
Meiju Liu, Shuai Zhang, Kaiwen Duan · 2024
Image anomaly detection is a significant application in the context of large-scale industrial manufacturing. To enhance the performance of image anomaly detection, we propose a superior approach called the reverse information distillation (RID) model, which is based on the reverse knowledge distillation (RD) model. Within the knowledge distillation process, we employ informative knowledge distillation (IKD) to extract informative knowledge and provide a strong supervisory signal to mitigate overfitting issues. Furthermore, we introduce an attention module in the fusion and compression of features, resulting in the creation of an attention bottleneck embedding (ABE) module that mitigates the loss of feature information. Experimental results on the public MVTec AD dataset demonstrate that RID achieves an AUROC of 98.9% in anomaly detection, with corresponding AUROC of 98.2% and PRO of 94.9% in anomaly localization. The effectiveness of our method is demonstrated by its superior performance compared to RD.