Industrial anomaly segmentation algorithm based on asymmetric teacher-student reconstruction network

Industrial anomaly detection is crucial in industrial production. Unsupervised methods based on teacher-student distillation networks have been widely used in this field. However, most of the existing teacher-student distillation networks use the similar or identical neural network structure, which hinders the diversity of anomaly feature and leads to a decrease in the accuracy of anomaly segmentation. To address this issue, this paper proposes an industrial anomaly segmentation algorithm based on an asymmetric teacher-student reconstruction network. The algorithm consists of a pretrained teacher network, a student reconstruction network, and a segmentation network. The teacher network is constructed with ResNet18 and the student network uses an encoder-decoder architecture. To better reconstruct features matching the teacher network, a one-class embedding module based on multi-scale feature fusion is introduced into the student reconstruction network. It effectively retains the multi-scale features in the encoder and compensates for the main information loss in the decoder. Additionally, in the segmentation network, we introduce the spatial-channel reconstruction convolution module, which reduces the redundant information between features and effectively improves the performance of the algorithm. Finally, the experimental results on the public dataset MVTec AD show that our algorithm has better accuracy compared with current mainstream algorithms. The average accuracy on image level and pixel level AUROC reached 99.2% and 99.0%, respectively, proving the effectiveness of our method.