Yolov4 High-Speed Train Wheelset Tread Defect Detection System Based on Multiscale Feature Fusion

Table 1

Feature extraction backbone network.

Block name	Type	Filter size	Output size
CBM	Conv1	3 × 3, 32, stride = 1	416  ×  416
Csp1(F1)	Conv2	3 × 3, 64, stride = 2	208 × 208
	RS1		208 × 208
Csp2(F2)	Conv3	3 × , 128, stride = 2	104 × 104
	RS2		104 × 104
Csp3(F3)	Conv4	3 × 3, 256, stride = 2	52 × 52
	RS3		52 × 52
Csp4(F4)	Conv5	3 × 3, 512, stride = 2	26 × 26
	RS4		26 × 26
Csp5(F5)	Conv6	3 × 3, 1024, stride = 2	13 × 13
	RS5		13 × 13
CBL	Conv7		13 × 13
SPP	MaxPool		13 × 13
CBL(F6)	Conv8		13 × 13