Unsupervised industry anomaly detection via asymmetric reverse distillation

Abstract

Existing unsupervised industry anomaly detection methods often rely on convolutional operations to capture fine-grained details in images. However, they may overlook crucial global context embeddings that are essential for accurate industry anomaly detection. To tackle this issue, we propose a global attention module, known as Global Attention with Spatial location and Content (GASC), which extracts global embeddings by considering both spatial information and content, thus compensating for the limitations of convolutional operations. Moreover, we introduce a novel residual unit with GASC and pre-activation in the student network, resulting in an asymmetric reverse distillation network (ARD). This architecture addresses the problem faced by previous methods where teacher and student networks share identical or similar structures, making it challenging to extract distinctive features for industry anomaly detection. Furthermore, to enable ARD to detect anomalies of various sizes, we incorporate both local details and global semantics by comparing the discrepancies between teacher and student network embeddings using cosine similarity at multiple scales. Finally, our approach is extensively evaluated through quantitative and qualitative experiments conducted on the MVTec, BTAD and KolektorSDD2 datasets, showcasing the outstanding anomaly detection performance and generalizability of our method.