Lightweight detection method of coal gangue based on multispectral and improved YOLOv5s

Aiming at the problems of poor identification effect and low separation efficiency of existing coal gangue identification and separation methods, this paper puts forward a lighter coal gangue identification and detection scheme based on YOLOv5s (You Only Look Once Version-5s) and multispectral image technology. Set up a multispectral acquisition system and shoot the required data sets. Three bands with high recognition accuracy and low correlation are selected from 25 bands in the spectrum to form RGB images for model detection. A depth separable convolution is introduced into the feature fusion network, which makes the model lighter, better fuses multiscale feature information, adds attention mechanism to the network, strengthens the objective attention, improves the ability of dense object detection and antibackground interference, reduces the complexity of the model, and improves the detection accuracy. The experimental results show that the improved YOLOv5s-MobileNet model has a mean average precision (mAP) of 98.88% on the coal and gangue test set. Compared with YOLOv5s algorithm, the model size is reduced by 12.5%, and the mAP is increased by 12.88%, which is beneficial to edge deployment to the greatest extent.