RDMS: Reverse Distillation with Multiple Students of Different Scales for Anomaly Detection

Unsupervised anomaly detection is a critical task in computer vision, often approached as a one-class classification problem. Knowledge distillation has shown promising results in this area, particularly with the emergence of reverse distillation networks using encoder-decoder architectures, which further enhance anomaly detection accuracy. In this study, we propose a novel reverse knowledge distillation network with multiple scale student decoders, called RDMS. RDMS integrates a pretrained teacher encoding module, a trainable multi-level feature fusion connection module (MFFCM), and a student decoding module composed of three mutually independent decoders. Each student decoder is tasked with distilling a specific feature from the teacher encoder, effectively mitigating the overfitting issue that occurs when the student and teacher structures are similar or identical. Our model achieves an average of 99.3% image-level AUROC and 98.34% pixel-level AUROC on the publicly available dataset MVTec-AD, and also achieves state-of-the-art performance on the more challenging BTAD dataset. The proposed RDMS model demonstrates high accuracy for anomaly detection and local-ization, highlighting the potential of multi-student reverse distillation for improving unsupervised anomaly detection capabilities. Source code is available at https://github.com/zihengchen777/RDMS