Adversarial Neighbor Perception Network with Feature Distillation for Anomaly Detection

Anomaly detection has become a research hotspot in the field of intelligent manufacturing, which has attracted strong attention from academia and industry. Although the unsupervised methods based on reconstruction have shown promising results in anomaly detection, they still have the problems that the representation of abnormal features is not significantly different and the representation of normal features is not accurate enough. To solve these problems, we propose an adversarial neighbor perception network with feature distillation (ANP-FD) for anomaly detection, which includes multi-scale neighbor perception, robust distillation recovery, and co-attention adversarial detection modules. First, the multi-scale features pass through the neighbor awareness to predict accurate spatial location information to improve the reconstruction accuracy. Then, the abnormal features are eliminated by a robust trainable distillation structure, which expands the representation difference of abnormal features during reconstruction. In addition, the co-attention adversarial detection can accurately detect and locate anomalies in the multi-scale feature space. The experimental results on the MVTec, BTAD, MNIST, Fashion-MNIST, and CIFAR-10 datasets demonstrate that our proposed method achieves better performance than current state-of-the-art (SOTA) approaches. Especially, the proposed method achieves 99.5 AUROC% and 94.5 AUPRO% on the MVTec. We also achieve superior performance in few-shot scenarios. Code: https://github.com/thesusu/ANP-FD.