Abstract: Aiming at the complex situation of dust,metal reflection,uneven illumination and inconsistent target scale in
mine smoke detection,which leads to slow detection speed and low accuracy,a mine smoke detection model YOLO-MSD based
on improved YOLOv8 n is proposed. Firstly,RepVGGBlock is introduced into the backbone network to improve the detection
performance under complex interference. Secondly,a shared convolutional pyramid (FPSC) is designed to replace SPPF,which
effectively improves the multi-scale feature perception ability and reduces the amount of calculation. Finally,the weighted fusion
idea of BiFPN is used to improve the multi-branch auxiliary feature pyramid network (MAFPN) as the network structure.
By designing shallow auxiliary weighted fusion (SAWF) and deep auxiliary weighted fusion (AAWF) modules,the accuracy of
small target smoke detection is improved and the number of parameters is effectively reduced. The experimental results show
that YOLO-MSD achieves 91. 7% AP50 and 74. 5% AP50-95 while maintaining the reasoning speed of 255 FPS,which is 4. 0
and 6. 1 percent points higher than that of YOLOv8n,respectively. It can still maintain stable smoke detection performance in
complex mine environment.

Key words: mine smoke detection,RepVGGBlock,YOLOv8n,multi-scale feature fusion,MAFPN