Joint decision normalizing flow based on frequency separation for industrial anomaly detection and localization
Xingzhao Hua, Yu Chen, Yueming Hu, Yili Chen, Zhenting Yan, Jun Han, HuiLing Huang · Neurocomputing · 2025
In industrial manufacturing processes, the detection of defective products caused by unforeseen circumstances is crucial. Although unsupervised anomaly detection methods have been widely applied in this domain, these methods often suffer from frequency bias during training, leading to uneven learning of frequency information by the model and resulting in missed or false detections. To address this issue, this study proposes a Joint Decision Normalizing Flow based on Frequency Separation. Specifically, Class Attention in Image Transformer(CaiT) is employed to separate low-frequency and high-frequency features of normal samples, respectively, while Normalizing Flow is employed to estimate the distributions in different frequency domains. Furthermore, a Feature Fusion Attention mechanism is proposed to effectively integrate high and low-frequency information, ensuring a balanced representation of detailed and structural features. Finally, decisions are derived by synthesizing the distribution information from these distinct frequency domains. This method mitigates frequency bias during training to enhance detection accuracy. Comprehensive experiments conducted on the MVTec AD dataset demonstrate that our method achieves an AUROC of 99.5 % for image-level anomaly detection, representing state-of-the-art performance with superior robustness. Additionally, an AUPRO of 98.2 % validates the method’s effectiveness in detecting fine-grained anomalies.