EdgeAD: Unsupervised Learning Model Based on Prior Knowledge Enhanced Image Anomaly Detection of Heavy Railway Freight Cars
Weiyu Zhang, Hongmei Shi, Ji Wei Qiu, Zujun Yu, Jianbo Li · IEEE Transactions on Instrumentation and Measurement · 2025
Trouble of moving freight car detection system (TFDS) is an important safeguard measure for the operation of railway freight cars. For few- and zero-shot faults that seriously affect traffic safety, it is a tough problem to realize automatic TFDS image recognition. In recent years, anomaly detection (AD) models based on unsupervised learning have provided a solution to this problem, but the current accuracy (ACC) and detection efficiency still need to be improved. The current AD model mainly reconstructs the image with the global information of the whole image and lacks the effective employment of focused information. This article proposes an AD algorithm with prior knowledge to enhance the ACC and reduce the time cost. First, an edge information-guided autoencoder image reconstruction algorithm is designed using edge information of TFDS images as the prior knowledge, which achieves improved performance by eliminating invalid information. Second, an edge-weighted loss is proposed to enhance the constraining effect of edge information during the training process. Finally, an empirical mask mechanism is proposed, which is guided by the prior knowledge of the anomaly characteristics. Validated on the real-world TFDS-based dataset, the proposed model improves the detection performance by 21.8% compared to the variational autoencoder (VAE) model. The proposed model is also extensively and comprehensively quantitatively evaluated against mainstream industrial image AD models, and the proposed model has better ACC and detection speed in TFDS image AD.