Defect Detection of Wheel Set Tread Based on Improved YOLOv5

Fig. 6. Comparison of detection results between YOLOv5 and improved YOLOv5. (a) (c) (e) Detection results of YOLOv5; (b) (d) (f) detection results of the improved YOLOv5

Download full size

Fig. 7. Comparison of inclusion detection results. (a) Scratch; (b) including scratch; (c) confluent scratch

Download full size

Table 1. Target quantity in data
View table
Table 1. Target quantity in data
Defect type Training set Validation set Test set Total
Scratch 2074 254 266 2594
Laceration injury 1037 143 130 1310
Peel 1425 174 183 1782

Table 2. Experimental results of different improvements
View table
Table 2. Experimental results of different improvements
Model CBAM Reduce layer EIoU Precision /% mAP /% Speed /（frame·s^-1）
YOLOv5 86.3 87.1 45.3
Improvement 1 + 89.2 91.6 44.1
Improvement 2 + + 88.5 90.8 48.1
Improvement 3 + + + 90.7 92.6 48.1

Table 3. Comparison of model complexity
View table
Table 3. Comparison of model complexity
Method Model size /MB Parameters /10⁶ FLOPs /10⁹
YOLOv5s 13.6 7.03 15.9
Improvement 1 14.0 7.04 16.2
Improvement 2 10.3 5.28 15.5
Improvement 3 10.3 5.28 15.5

Table 4. Performance comparison of mainstream target detection algorithms

View table

Table 4. Performance comparison of mainstream target detection algorithms

Model		AP（R_IoU=0.5）/%		mAP/%	Speed /（frame·s^-1）
Model	Tread scratch	Tread bruise	Tread peel	mAP/%	Speed /（frame·s^-1）
SSD	72.4	67.9	71.9	70.8	55.4
Fast-RCNN	89.3	88.2	87.2	88.2	2.2
YOLOv5	87.1	85.2	89.1	87.1	45.3
YOLOv5s_A	88.8	88.7	89.9	89.2	45.1
YOLOv5s_B	85.9	84.2	88.6	86.2	52.3
Proposed model	92.1	92.3	93.3	92.6	48.1