Abstract: Industrial image anomaly detection is a critical component in large-scale industrial manufacturing.Addressing challenges such as difficulty in annotating anomalous samples and obtaining prior information about anomalous regions in industrial image anomaly detection,a model based on asymmetric teacher-student networks for unsupervised image anomaly detection is proposed.Firstly,to tackle the problem of over-imitation mapping caused by high similarity in structure between teacher and student networks,an asymmetric teacher-student network is designed.Contextual Transformer modules are introduced into the residual blocks of the student network to add structural diversity to the teacher-student networks,preventing the student network from over-imitating the mapping of the teacher network.Secondly,to enhance the generalization difference between teacher and student networks,a moving average normalization layer is introduced into the teacher network to improve detection performance.Finally,a multi-scale abnormality map fusion mechanism is introduced to better detect anomalies of different sizes by fusing abnormality score maps of different scales.Experiments conducted on the MVTec AD public dataset show that the proposed method achieves an image-level AUROC of 95.7% and a pixel-level AUROC of 97.4%,verifying the feasibility and effectiveness of the approach.

Key words: Anomaly detection, Knowledge distillation, Transformer, Unsupervised learning, Multi-scale features