Abstract: Paste image recognition is an effective method for monitoring paste quality. Proposes a paste image recognition method based on convolution neural network, which can realize high-precision recognition of paste state. By collecting the images of the tailings suspension in three states of slurry, paste and filter cake, through image preprocessing and data set division, combined with the method of transfer learning, AlexNet model, VGG16 model, VGG19 model and ResNet50 model of the convolution neural network are pre-trained, the recognition accuracy and loss value of the four models are compared to determine the best model. The model is optimized using Adam algorithm and RAdam algorithm, and the recognition results of two optimizers are compared. The results show that the four classical convolution neural network models have good performance in paste image recognition, and ResNet50 model has the best performance. Compared with Adam algorithm, RAdam algorithm optimization algorithm had faster convergence speed and higher recognition accuracy. The recognition accuracy of paste image based on RAdam algorithm convolution neural network can reach 99.24%, which can realize high-precision recognition of paste image.