Dual‐task teacher–student networks for unsupervised anomaly detection

Abstract In recent years, unsupervised anomaly detection methods have garnered significant attention in colour textile anomaly detection. However, most current methods lack constraints from anomalous samples during training. Even when anomalies are present, such as in distillation networks, the student network is highly likely to overgeneralise, limiting the performance of anomaly detection. To address this problem, this study proposes a dual‐task teacher–student unsupervised anomaly detection (DTTS) model for colour textile anomaly detection. DTTS improves the robustness of the model to normal data by dual‐tasking knowledge distillation and reconstruction, so that the student encoder shares the underlying representation. In addition, DTTS introduces a mask convolution strategy to enhance data diversity and model stability. Experiments were conducted on the colour textile yarn‐dyed fabric image dataset. DTTS mitigates the overfitting problem of the student network during training and achieves better anomaly localisation performance. The anomaly detection achieves an excellent performance of 99.2% on the image‐level area under the receiver operating characteristic curve (I‐AUC), and the anomaly localisation achieves a performance level of 99.6% and 94.4% on the pixel‐level AUROC (P‐AUC) and the area under the per‐region overlap curve (AU‐PRO), respectively.