Multi-level Feature Fusion Guided Distillation for Industrial Anomaly Detection

Knowledge distillation has yielded promising results in the demanding task of unsupervised anomaly detection. The representation divergence of anomalies within the teacher-student (T-S) model plays a crucial role in anomaly detection. To enhance the model's capabilities, we designed a multi-level fusion-guided network for the student network to boost the representation capabilities of the shallow stages. We also introduced a low-level feature path with contextual information, to guide the feature extraction process and enhance the differentiation between the teacher and student models. A feature fusion block was designed for effectively merging features of different levels within the network. Experimental results have demonstrated that our approach outperforms previous methods in the field of anomaly detection.