RNaAE: A Novel Approach for Identifying Unseen Railway Anomalies

Detecting abnormal objects in railway track inspection images using vision-based technology is a crucial task for ensuring the safety of railway transportation. Traditional supervised object detection methods fail to achieve satisfactory results due to the diverse categories of abnormal objects and the lack of abnormal samples. Even though the widely used Autoencoder can leverage reconstruction errors to detect anomalies without using abnormal data, they tend to generate a relatively high number of false positives. In this paper, we address the task in an unsupervised manner and propose a novel Random Network-Assisted Autoencoder, called RNaAE, for identifying unseen abnormal objects. Specifically, we first design a learnable network to fit a randomly initialized stochastic network with fixed weights, where the difference between two predictions can then be used to estimate whether the candidate object is anomalous. After combined with a traditional Autoencoder, a Gaussian mixture model is then used to classify the candidate box into normal and abnormal by anomaly scores. Extensive experiments conducted on our collected railway anomaly dataset demonstrate that the proposed RNaAE exceeds previous stateof-the-art methods, achieving 98.23% and 92.02% in terms of AUROC and F1-score.