Two-stage reverse knowledge distillation incorporated and Self-Supervised Masking strategy for industrial anomaly detection

In recent years, unsupervised anomaly detection based on knowledge distillation has gained special attention and some promising results have been reported in the literature. However, there is still room to improve the sensitivity of the model to anomalies. To do so, in this paper, a novel two-stage training method in terms of reverse knowledge distillation is proposed for anomaly detection and localization. Firstly, self-supervised mask training is introduced after the initial training of reverse knowledge distillation, which contributes greatly to the model detection against random unknown anomalies by self-simulating anomalies and forcing repair so as to reinforce learning single-category prototype patterns. Then, with the aim to facilitate the anomaly localization, an anomaly feature diffusion module is employed, which strengthens the correlation between pixels and helps spread the anomaly information to the surrounding area by covering the central pixel and reconstructing the representation for features after diffused. Furthermore, inspired by the human memory mechanism, an innovative normalized embedding memory bank is adopted to regulate the low-dimensional representations after embedding the encoding, inhibit the flow of anomalous information to the student decoder, and encourage the high-quality reconstruction of the model. Finally, the contextual similarity loss is used to guide the student model to learn knowledge representations from a contextual perspective, capture higher-order similarities between teachers and students, and delicately evaluate the differences between teachers and students. The empirical experiments conducted on the MVTec dataset show that the proposed SSMRKD method can achieve the best performance compared to other state-of-the-art methods, meanwhile extensive experiments of the ablation study validate the contribution of each component of the model. In addition, the advanced performance achieved on four commonly used datasets verifies the generalizability of the model in the industrial domain. Overall, the proposed SSMRKD method has significant advantages over the state-of-the-art anomaly detection methods.