Crack Detection Method of CRTSⅡ Track Slab Based on Faster R-CNN Improvement

The accuracy and timeliness of ballastless track damage detection and maintenance are crucial to the operation safety of high-speed railway, and the use of machine vision technology for crack damage detection of ballastless track slab on high-speed railway can greatly improve the accuracy and efficiency of detection work. Therefore, according to the sample data characteristics of CRTSⅡ track slab crack damage, a method based on Faster R-CNN improvement is proposed for track slab crack detection. The improved method transforms the detection issue into a positioning one, streamlines the network model, and selects the residual network as the backbone to avoid the decrease in learning speed due to the excessive depth of the network. Then, this method introduces a guide anchor frame to reduce the number of redundant anchor frames, thus improving the detection pertinence. Finally, it uses the Soft-NMS algorithm to improve the overlapping issue in track slab crack detection and improve its effects. In order to evaluate the reliability of the improved method, the evaluation criteria for CRTSⅡ track slab crack detection are established, according to which, the improved method is used to conduct comparison tests for comparing it with R-FCN, YOLO-v5, Faster R-CNN and YOLOx network algorithms. The results show that, with higher accuracy and minimal miss detection rate, the proposed improved method outperforms other algorithms in terms of overall performance. The precision of the optimal model reaches 95. 9%, and the recall rate is 89. 6%, which are improved by about 2%-4% and 2%-6%, respectively, compared with other classic algorithms, and can be well applied to CRTS Ⅱ track slab crack detection scenarios. © 2023 Chinese Academy of Railway Sciences. All rights reserved.