Railroad Track Defect Detections using Deep Learning

However, railroad transportation continues to be one of the most important components to global terms of freight and passenger transport, and is integral to the success of economic development and logistics efficiency. With bigger demand for rail service comes ever more need to not only maintain functionality of the rail system, but also to keep it safe and reliable. With railroad tracks being the case, one of the most pressing concerns at hand is to be able to detect defects in the tracks early and accurate. Due to labor intensive, time consuming and human error prone nature of traditional inspection methods, there is an immediate requirement of approaches to automated, intelligent inspection. Using high resolution inspection images of railroad tracks, this study investigates the use of deep learning techniques for automation in detecting defects in railroad tracks. Then based on the leakage of convolutional neural networks (CNNs), transfer learning strategies and real time object detection frameworks (e.g., YOLOv5), provide the good enough track anomaly classification and locating capabilities such as cracks, broken rail, loose fasteners and vegetation interference. Experimental results show that the classification accuracy was 94% for InceptionV3 model, and the real time detection using YOLOv5 has the mAP of 0.89 and the inference time of just 23ms per image. Deep learning appears to be a robust scalable and practical method of improving railway safety and maintenance operation.