Symmetry-Driven Unsupervised Abnormal Object Detection for Railway Inspection
Taocun Yang, Yuming Liu, Yaping Huang, Junbo Liu, Shengchun Wang
Abstract
Vision-based abnormal object detection in railway track inspection images is one of the critical tasks to ensure the safety of railway transportation. Even though many machine learning-based methods have been developed, these approaches heavily rely on anomaly supervisions and therefore cannot detect unknown anomaly classes. In order to tackle the problem, this article proposes a novel unsupervised method to detect abnormal objects, which does not require abnormal object training data. Specially, we find that a railway track image is almost symmetrical about the track central line, i.e., normal objects appear repeatedly and symmetrically, while abnormal ones are rare and also significantly different from the corresponding symmetric areas. Motivated by this observation, we propose to train a metric-learning-based deep model to learn the similarity between normal objects and the corresponding symmetrical areas. Then for each object proposal in one test image, we measure the distance between the proposal and the corresponding symmetrical regions, and determine whether it is an abnormal object based on the symmetrical metric. Extensive experiments on our collected dataset show that our proposed method achieves competitive performance compared with the state-of-the-art methods.