Anomaly Detection Method for Railway Infrastructure Based on UAV Imagery

The safety of railway infrastructure heavily relies on timely and accurate defect detection in key components such as bolts, insulators, fasteners, and noise barriers. This paper proposes an enhanced anomaly detection framework based on the PyramidFlow model, augmented with an Anomalous Sample Synthesis Module and a Dual Attention (DA) mechanism. The proposed approach is validated on a custom dataset captured by a DJI Matrice 600 UAV equipped with a Z30 zoom camera, encompassing both normal and defective samples across four categories. Experimental results demonstrate that the synthesized anomalous samples effectively improve feature discriminability, particularly in texture-rich categories such as bolts and insulators. Furthermore, the integration of the DA module significantly enhances the model's capability to detect small-scale defects and suppress background noise, leading to consistent improvements in both pixel-level and image-level AUROC across all categories. Notably, our DS-PyramidFlow model achieves an image-level AUROC of 100% and pixel-level AUROC of 95.58% in bolt detection, outperforming baseline methods. These findings highlight the potential of multi-level feature alignment and synthetic defect generation in robust anomaly detection for UAV-based railway infrastructure inspections.