Defect classification based on deep features for railway tracks in sustainable transportation

Abstract

Abstract Rail tracks are the most important component of train movement in rail transportation. Therefore, real-time detection of defects on track surfaces is important but also difficult because of the noise, low contrast, and inhomogeneity of density. In recent years, tools have been developed for robust and highly accurate defect detection with advances in deep learning technologies. However, the existing deep learning algorithms require a large number of parameters to be set, which is computationally expensive. Therefore, those algorithms cannot fulfill the requirements for quick inspection. In this study, rail surface defects were detected by fusing the features of two deep learning models. SqueezeNet and MobileNetV2, the two models selected for this purpose, are both smaller in size and faster than other deep learning models. However, both of these models are less accurate than other models. Therefore, in this study, a fusion model with high accuracy is proposed by combining the features of the two models. First, a contrast adjustment is applied to the original image of the rail, and then the rail track location is determined. Then, most weighted features are selected from each network, and the defects are determined by giving the reduced features to Support Vector Machines (SVM). Experimental results show that the proposed method gives better results for multiple rail surface defects under low contrast than using a single deep learning model.