A New Mobile Segmentation Network Approach for Defect Detection of Rail Surface

Railways have been carrying people and their loads for over 200 years. Rail inspections must be carried out periodically for safe transportation on railways. The periodic inspections on high-speed train lines prevent serious accidents. The control of the rails is carried out manually by the personnel of the railway enterprises. Rail inspection with computerized systems will reduce the time spent for control, minimize the errors that may arise from human perception, and reduce the cost. In this study, a hybrid approach using MobilenetV2 in the encoder part of UNET is presented to detect rail surface defects. The proposed approach has been applied to a special dataset prepared by the authors and a benchmark dataset, Rail Surface Discrete Defects (RSDD-I). According to the dice score metric, the hybrid model achieved 0.8702 success. According to the results of the study, the hybrid MobileNetV2 network with fewer parameters, faster training, and prediction time can be preferred for segmentation tasks on devices with low hardware power.