Data Enhancement Based on Adding Noise to Improve The Detection of Rail Surface Defects and Sleeper Fractures

Railway track inspection and maintenance is a significant topic. This research introduces a track inspection method based on image recognition to attain a rapid and effective track inspection methodology. The approach employs image enhancement technology to augment the dataset’s content, thereby improving the detection accuracy of the track detection model. The experiment indicates an improvement of substantial accuracy when employing image enhancement technology to augment the dataset, irrespective of the YOLOv5n or YOLOv5s model. The model’s average accuracy (mAP50) in detecting sleeper breaks reached the highest score of 99.5%. The datasets are publicly available at https://universe.roboflow.com/softteacherrail-xxepi/l10_ul_50_2.