Coal gangue target detection algorithm based on MBP-YOLOv8

In order to solve the problems of high redundancy of model parameters and difficulty in recognizing small target gangue in deep learning based coal gangue sorting method. A lightweight detection model (MBP-YOLOv8) for coal gangue sorting is proposed. First, the lightweight network MobileNet V3 is used as the feature extraction network, which accelerates the running speed of the model and effectively reduces the redundancy of model parameters. Second, a bidirectional routing attention mechanism (BiLevelRoutingAttention, BRA) is added at the end of the model backbone, which enhances the ability of the model to capture the details of the gangue texture and improves the detection quality. Finally, Wise-IoU V3 with adjustable weight coefficients is used as the regression loss, which improves the convergence speed and robustness of the model as a whole. The experimental results show that the mAP of the proposed method is 90.3%, the number of parameters is reduced by 21.5%, and the frame rate is 138 frames/second. It effectively improves the sorting accuracy of gangue while satisfying the real-time detection of gangue.