A Real-Time Application for Rail Surface Defect Inspection Utilizing Rectangular-Shaped Labels

During the operation of high-intensity trains, various types of defects often arise, resulting in minor to moderate damage to the rail surface. Surface anomalies on railroad tracks can lead to increased speeds, resulting in elevated noise levels and a higher risk of train accidents. To enhance quality standards, manual inspections by field workers are necessary. However, these inspections require significant manpower, suffer from accuracy issues, and incur substantial costs. To streamline the inspection process, we analyzed a deep learning-based surface flaw detection system that employed three variations of the You Only Look Once (YOLO) algorithm: YOLOv6, YOLOv7, and YOLOv8. The aim was to improve the efficiency of the sorting stage. Furthermore, our experiments focused on converting pixel labels into rectangular or bounding box labels using the RSDDs dataset, which comprises two primary categories: high-speed rail (type 1) and heavy rail (type 2). Given the challenging nature of this dataset, the defect detection system achieved accuracies of 92.7% for YOLOv6-L6, 95.6% for YOLOv7-D6, and 99.5% for YOLOv8-S within the type 1 category. In the type 2 category, the results were 88.03% for YOLOv6-S6, 88.5% for YOLOv7W6, and 91.3% for YOLOv8M. These comprehensive experimental findings demonstrate that the YOLOv8 variant holds great potential in terms of mean average precision (mAP) accuracy for rail surface inspection systems utilizing rectangular-shaped labels.