Detection and classification of rail surface defects using image processing and artificial intelligence methods

Railways have evolved from the invention of the steam engine to the high-speed trains of today. Since the first day, locomotives and wagons have moved on iron rails. Rails are critical for railway vehicles. Defects are formed on the rail surfaces over time. These can be rolling stock-based or production-based. In this study, squat and cold bruising defects were classified among the defects on the rail surfaces. Three data sets were prepared using images from the Turkish State Railways (TCDD) Research Center Directorate. The first data set (VS-1) contains 447 images. The second data set (VS-2) is the CLAHE filter applied on VS-1. The third data set (VS-3) was obtained by augmenting data on VS-2. For classification, we used our own model and learning transfer models VGG-16, ResNet50 and DenseNet121. The models were trained on all data sets. VS-1 VGG-16 model achieved 80% accuracy. With VS-2, 95% accuracy was achieved in the DenseNet121 model. With VS-3, 95% accuracy was achieved in the ResNet50 model. Pre-processing and data augmentation of the samples in the dataset had positive effects.