Real-Time Object Detection Algorithm for Railway Crack Inspection System Comparison of Yolo V6, V7, V8 and Nas

The increasing intensity of train usage in big cities has resulted in some damage to the railroad tracks, one of the defects that appear is on the surface of the railroad, and this maintenance will require effort from workers every time to check and maintain the surface of the railroad periodically. This will have an impact on the need for labor if done manually for periodic inspections, and of course, will increase the need for operational costs and others. To streamline time efficiency in field inspections we analyze and compare surface defect detection systems based on deep learning using several variations of the You Only Look Once Algorithm (YOLO): YOLOv6, YOLOv7, YOLOv8, and YOLO NAS. In addition, our experiments focus on the training results using the RSSD dataset consisting of highspeed rail (type1) and heavy rail (type2), We aim to assess the effectiveness of each model variation in detecting defects on the surface of the railroad tracks and performance matrices such as recall, precision, F1-Score and Mean Average Precision are used. In this research, the recall matrix is needed to minimize the false negative value to detect defects, so that the error is small when determining the point that must be repaired immediately. In accordance with the existing challenges from the Yolo v7, v8, and NAS data sets, they are good at remembering damaged objects during training and validation. While for Yolov6 it is low for the recall matrix but good for the precision matrix.