Negative Bidirectional Distillation Model based on Self-Supervised Learning

Knowledge distillation techniques have shown significant potential in the challenging field of unsupervised anomaly detection (AD). The difference in anomaly representation in teacher-student (T-S) models provides crucial insights for AD research, but several core issues remain in practical applications. First, the limited samples make it difficult for traditional techniques to learn abundant, generalizable representations. Second, due to the lack of direct involvement of anomaly samples, it’s difficult for the model to learn discriminative features that effectively distinguish between normal and abnormal samples. Furthermore, using similar or identical architectures in T-S models might limit the diversity of anomaly representation, making it difficult for the model to build a stable decision boundary to accurately distinguish anomalies. To address these challenges, we propose the NBDM (Negative Bidirectional Distillation Model) method, which adopts a bidirectional distillation architecture combined with self-supervised learning and simulated anomaly strategies to distill the pre-trained teacher network’s feature extraction capability on normal samples. Meanwhile, NBDM introduces a negative learning mechanism that restricts the model’s handling of anomalies, reducing the adverse impact of anomaly features on model performance. Experimental results on the MVTec unsupervised anomaly detection benchmark dataset show that NBDM achieves significant performance improvement over existing methods, demonstrating stronger competitiveness and advantages.