Comparative Study of Classification and Detection-Based Approaches for Railway Track Fault Identification Using Deep Learning

Abstract

Railway track inspection is still commonly carried out through manual checking, which takes considerable time and can be affected by human error. This work compares two deep-learning approaches for identifying faults in railway-track images: image classification and object detection. A ResNet18 model is used first to classify an image as Fault or No Fault. Although this gives a useful image-level result, it does not indicate the position of the damaged region. To overcome that limitation, a YOLOv8 detector is trained to mark faulty areas with bounding boxes. The detector is evaluated on the railway-track dataset and then tested on a larger structural-crack dataset to examine its ability to work with related visual defects. The experiments show that detection is more useful for practical inspection because it provides both fault identification and location. The results also indicate that additional training data improves the usefulness of the detection model.