Rail surface anomaly detection : a deep learning approach for computer vision

Abstract

Rail surface defects have become more of an issue in recent years due to new manufacturing techniques which produce head-hardened rails and as industry demands higher speeds, heavier loads and increased tra c. These defects can cause catastrophic accidents, which have consequences such as death, injury, huge cost implications and loss of public con dence. Computer vision systems have become popular, as cameras are non-contact full- eld sensors which are low in cost, have high sampling rates and provide appealing performance. However, accurate inspection remains challenging due to dynamic non-linear environmental and rail surface conditions in which images are captured, which result in a heterogeneous image dataset. It is also di cult to select useful features which satisfy the variations due to di erent failure modes. In addition, there is a class imbalance issue, as most captured images do not contain any defects. In this dissertation, we develop deep generative models that are trained exclusively using healthy images of a rail surface so that we learn useful features to capture the complex nature of the images which are acquired. We propose multiple models which operate with images at di erent resolutions. We present a new dataset which will be made publicly available. Experimental results demonstrate that our proposed models can perform accurate detection using our dataset. The proposed algorithms are highly parallel and computationally e cient, which enables real-time inspection at speeds that exceed the world's fastest railway trains: Fuxing Hao CR400AF/BF that has a continuous operation speed of approximately 400 km=h.