RAILWAY DEFECT DETECTION USING MACHINE LEARNING

Abstract

As railway tracks are used over time, they can develop permanent defects on their surfaces. These defects can get worse quickly and pose dangers to train safety. Therefore, it's important to detect these rail defects accurately and quickly. However, this task is challenging because the defects don't always stand out clearly from the track, vary greatly in size, and there aren't many examples to train detection systems on. To overcome these issues, we propose using YOLOv5 to detect rail surface defects. We first improved our dataset of rail defect images by applying techniques such as flipping the images, randomly cropping them, and adjusting their brightness. We also incorporated pretrained model classification methods like EfficientNet and ResNet-50 to reduce the network's computational load. Our tests indicate that the improved YOLOv5 algorithm achieves a 96.9% average precision in detecting rail surface defects, proving it to be effective and efficient for practical use in engineering.