Coal gangue detection method based on improved YOLOv5

Accurate identification of coal and gangue is helpful to improve the quality of raw coal. Based on the YOLOv5 algorithm, a new model yoloV5-SEDE with high precision and fast convergence is proposed. First, Mosaic data enhancement method is used to randomly crop and scale images to form new images to expand the data set. Nine anchor frames are generated by k-means clustering. After that, SENet attention module was added to enhance the identification ability of coal gangue targets, suppress invalid features and improve the utilization rate of feature information. SPPF structure is used instead of SPP structure to enrich the expression ability of feature mapping. In the feature fusion stage, the original feature pyramid fusion method is replaced by BiFPN weighted feature pyramid fusion method to make the semantic features obtained from the network have stronger robustness. Compared with the original target recognition algorithm, the average accuracy of this algorithm is improved by 1.59% (MAP_0.5) and 2.57% (MAP_0.5:0.95) respectively.