Jointing High-State and Low-State Feature Reconstruction Networks and Discriminator Training for Track Anomaly Detection
Ziyi Wang, Yun Ye, Hongxing Xie, Xingke Zheng, Liyan Zhang, Enguo Chen, Sheng Xu, Tailiang Guo · 2024
Railway safety inspections are critical to prevent catastrophic incidents, yet existing anomaly detection methods often suffer from insufficient sample data and overgeneralization. This paper proposes an improved anomaly detection method called HLFRN-D, which integrates dual-state feature reconstruction networks trained alongside a discriminator. HLFRN-D mitigates decoder overgeneralization by shifting the discriminator's focus from pixel-level abnormalities to evaluating the reconstruction performance of the networks. An anomaly generation module and a novel loss function further enhance robustness. Experimental results show that HLFRN-D achieves superior performance, with an AUROC score of 97.6% and PRO score of 94.4% on MVTec AD, and 99.3% AUROC and 97.7% PRO on track datasets, highlighting its effectiveness in detecting diverse anomaly types and locations.