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

Algorithm 1

Multiscale feature fusion.

(i)	Input: Feature map {F2, F3, F4, F6} generated from backbone network
(ii)	Preprocessing: Normalize {F2, F3, F4, F6} with L2 normalization
(iii)	While epoch >0 and input feature map is not empty do
(iv)	Downsample feature map F2 to 52 × 52 as F2′ through 3 × 3 conv
(v)	Concate F2′ and F3 as FF1′
(vi)	Get FF1 through 1 × 1 conv
(vii)	Downsample feature map F2′ to 26 × 26 as F2″ through 3 × 3 conv
(viii)	Concate F2″ and F4 as FF2′
(ix)	Get FF2 through 1 × 1 conv
(x)	Downsample feature map F2″ to 13 × 13 as F2‴ through 3 × 3 conv
(xi)	Concate F2‴ and F6 as FF3′
(xii)	Get FF3 through 1 × 1 conv
(xiii)	End
(xiv)	Output: Fusion feature {FF1, FF2, FF3}