In order to improve the degree of automation of separating coal and gangue, the coal and gangue original image collected on site is used as input to establish a coal and gangue image recognition model based on convolutional neural network in this paper to solve the image feature values need to be manually selected and the robustness is poor. The convolutional neural network is visualized by deconvolution, and the process of extracting coal and gangue image features by convolutional neural network is analyzed. Decomposing the original image by biorthogonal wavelet, the wavelet transform layer is set up in the convolutional neural network. The convolution operation is performed by combining the high frequency coefficient with the original image to optimize the recognition effect of the model. The results show that this model can effectively differentiate the coal and gangue images and has strong generalization ability. Setting the wavelet transform layer can improve the network training efficiency and recognition accuracy. When combining the second layer high-frequency coefficient of wavelet transform with the original image, the network model is optimal. Compared with the traditional recognition model, the model has better adaptability under different illumination conditions, and the recognition accuracy of the test set reaches 93%.