Aiming at the problems of complex process and low detection accuracy of artificially extracting coal gangue image features in traditional coal gangue detection algorithms,a light-weighted PAM-M-YOLO coal gangue detection model was proposed.Firstly,the MobileNetv3 feature extraction network was used to replace the original model backbone network,and the depth separable convolution was used to replace the traditional convolution to extract the features of coal gangue images.Secondly,PAM parallel attention module was designed to improve the attention of feature map channel and spatial information after the splicing of target detection network layer.Finally,a priori information is added to the model based on the CAM activation restriction branch to reduce the local collapse of the model on non-key features.The experimental results show that the accuracy,recall rate and mAP value of the light-weighted PAM-M-YOLO coal gangue detection model are 98.7％,97.5％and 98.8％,respectively,which are 3.6,2.3 and 2.0 percentage points higher than those of the original M-YOLO model.The number of parameters is 3.8 MB,which is nearly 1/2 lower than the YOLOv5 model.The visualization effect of the heat map shows that the information concerned by the light-weighted PAM-M-YOLO model in the detection process is more concentrated in the coal gangue area,which can effectively solve the local collapse problem of the model in the coal gangue area.
Coal gangue image detectionYOLOv5 modelLight-weighted PAM-M-YOLO modelDeep learningAttention mechanismLoss function
维普
万方数据
