Industrial Anomaly Detection via Teacher Student Network

This paper proposes a teacher-student network for industrial anomaly detection tasks using knowledge distillation techniques, where the feature representation capabilities of the teacher network model are utilized to guide the student model in efficiently identifying anomalous images. In detail, the VGG model is introduced as the backbone network for anomaly detection. Meanwhile, a residual attention mechanism is proposed to improve the feature output capability of the student network. The student network is able to learn good feature extraction abilities due to the robust feature representation of the teacher network. At the same time, the residual attention mechanism enhances the student network's ability to characterize the output of normal data features, amplifying the feature differences between the student and teacher networks for abnormal data. In summary, this paper addresses the problem of difficult detection of small targets in anomaly detection. Experimental validation has been performed on the industrial anomaly detection dataset MVTec-AD, which ultimately achieved 96.70% and 97.94% results for detecting and localizing, respectively. Compared with the other anomaly detection methods, the method proposed in this paper achieves a significant performance improvement.