Author Affiliations
1School of Mechanical Engineering, Sichuan University of Science & Engineering, Yibin, Sichuan 644000, China;2Chengdu Institute of Computer Application, Chinese Academy of Sciences, Chengdu, Sichuan 610041, China3School of Mechanical Engineering, Sichuan University, Chengdu, Sichuan 610065, Chinashow less
Fig. 1. Schematic of collection equipment
Fig. 2. Collected wheelset images
Fig. 3. Flow chart of dual network detection algorithm
Fig. 4. Areas with contrasting gray values
Fig. 5. Direction-corrected wheelset images
Fig. 6. SSD network structure
Fig. 7. YOLOv3 network structure
Fig. 8. Distribution chart between original Anchor Boxes and target size
Fig. 9. M-YOLOv3 network structure
Fig. 10. Distribution chart between new Anchor Boxes and target size
Fig. 11. P-R curves of three models when extracting tread
Fig. 12. Partial identification renderings of the three networks. (a) SSD300; (b) YOLOv3; (c) YOLOv3-tiny
Fig. 13. P-R curves of five models when identifying defects
Fig. 14. Partial identification renderings of five kinds of networks. (a) M-YOLOv3; (b) YOLOv3; (c) YOLOv3-tiny; (d) SSD300; (e) SSD512
Fig. 15. Missing and misdetection of some tread defects
Fig. 16. P-R curve of two detection methods
Network | AP /% | Time /ms | IOU /% |
---|
SSD300 | 99.8 | 31.8 | 83.2 | YOLOv3 | 96.9 | 44.3 | 80.1 | YOLOv3-tiny | 90.6 | 13.9 | 73.6 |
|
Table 1. Tread detection performance
Network | M-YOLOv3 | YOLOv3 | YOLOv3-tiny | SSD300 | SSD512 |
---|
AP /% | 89.9 | 90.5 | 72.2 | 19.3 | 42.0 | Time /ms | 58.7 | 63.2 | 17.5 | 24.5 | 46.4 |
|
Table 2. Defect detection performance
Network | AP /% | Time /ms |
---|
YOLOv3 | 73.2 | 92.7 | Exper1 & Exper2 | 89.7 | 92.3 |
|
Table 3. Defect detection performance of two detection methods