Developing defect detection algorithms by making meaning of rail and environment of rail with innovative deep learning methods

Accidents can happen on railways. Tracks need to be inspected to prevent accidents. In this study, "With deep learning methods, railway images can be interpreted and rail defects can be detected autonomously." explores hypotheses. Various image processing techniques were applied to the data set consisting of ray images collected by UAV. GoogleNet reached 96% accuracy and SqueezeNet 91.75% accuracy. Images in the Rail Surface Discrate Defect Type-1 dataset, which consists of publicly shared rail surface images, were sliced amplified and rescaled. UNet3+ achieved 81% accuracy at pixel level and 88% accuracy at defect level. With the help of the cameras placed under the locomotive, the collapse defects were estimated with the UNet, UNet3+ and hybrid MNV2UNet network with the data set. UNet and UNet3+ achieved 88% success, and the hybrid model 87%. In another study, Vgg16 and MobileNetV3-Small networks with customized classifier layers were trained. In the study in which the unsupervised learning approach was applied, the networks reached 98% and 96% accuracy rates, respectively. UNet, ICNet, BiSeNetV2 networks were trained with the dataset created by customizing the Railsem19 dataset, which is publicly shared, containing railway images. All three networks achieved an accuracy rate of 98%. Rail defects were detected with the first three data sets, and the railway images were interpreted with the last data set. According to these results, high-fidelity railway images can be interpreted and rail surface defects can be detected autonomously. The study can contribute to the rail inspections of railway enterprises.