AEKD: Unsupervised auto-encoder knowledge distillation for industrial anomaly detection

Unsupervised Anomaly Detection (UAD) has achieved promising results in industrial Surface Defect Detection. Knowledge-Distillation (KD) based UAD became a hotspot due to its simple structure and convincing detection results. However, the generalization issue of the similarity between Student (S) and Teacher (T) models in KD hinders the accuracy. KD based UAD is based on the feature differences between the T and S models, and the similar feature expressions of the T and S models would lead to the failure detection on the anomalous images. To cope with this issue, a new Unsupervised Auto-Encoder Knowledge Distillation (AEKD) is developed to accurately detect anomalies and the locate anomalous regions. AEKD uses the encoder as T model and the AE structure as S model. The structural differences between T-S models can effectively suppress the generalization issue. Firstly, a new S model structure is proposed to strengthen the structure difference of T-S model. Secondly, a trainable Multi-scale Features Fusion module is introduced to reduce anomaly disturbance. Thirdly, the different data flow of T and S model is designed to reinforce the different expression in T and S model to anomalies. AEKD has been conducted on the public MVTec, DAGM dataset and a real-world glass bottle dataset. The results validate that AEKD has achieved the excellent results by comparing with other famous UAD methods.