Dual-Task-Based Unsupervised Fabric Anomaly Detection

In recent years, the field of colored textile defect detection has seen a growing interest in unsupervised anomaly detection methods. However, most of the current approaches lack constraints on anomaly samples during the training phase. As a result, even when anomalies are present, student networks are highly susceptible to overgeneralization, which ultimately limits the performance of defect detection. To address this challenge, this study introduces an innovative dual-task Teacher-Student unsupervised anomaly detection model (DTTS) tailored for colored textile defect detection. The DTTS model effectively integrates two key tasks—knowledge distillation and reconstruction—to facilitate shared low-level feature representations in the student encoder, thereby significantly enhancing the model's robustness in handling normal data. Moreover, DTTS incorporates a masked convolution strategy, which not only enriches data diversity but also strengthens model stability, laying a solid foundation for efficient performance. Extensive experiments conducted on the YDFID colored textile dataset demonstrate that DTTS successfully mitigates the overfitting issue encountered by student networks during training and achieves impressive results in anomaly localization.