A fast recognition method for coal gangue image processing

This paper proposes a modified YOLOv4 model, named GYOLO, for coal gangue recognition with the aim of reducing model parameters, improving calculation speed, and reducing equipment requirements. To achieve this, the paper optimizes the feature extraction network structure by using linear operation instead of traditional convolution to obtain redundant feature maps, thus reducing the number of parameters by 29.7%. A feature fusion network structure is also reconstructed to strengthen the model's use of feature information, further explore the dependence of each channel feature, and make better use of feature information. The ablation experiment is designed to verify the effect of each improvement. The image is blurred to improve the difficulty of target detection and test the robustness of the GYOLO model. The generative adversarial network is trained with a small amount of coal gangue data, and then a large amount of virtual data is obtained by using the generative adversarial neural network. The GYOLO model is trained by transfer learning, which reduces the dependence of the model on real data. The GYOLO algorithm is compared with a variety of excellent target detection algorithms to analyze the performance of the algorithm. It is verified that the accuracy of the proposed method is 97.08%, which is 2.3% higher than that of the original model, the amount of parameters is reduced by 19.6%, and the amount of data required is reduced by 57.3%. The balance between data volume, parameter quantity and model performance is further realized.