Maps play an important role in people's production and life. There is a lot of information in annotations. Identifying map name annotation categories is of great significance for computer reading maps and further drawing maps in the future. Recently, popular deep learning technologies, especially convolutional neural networks, have a good effect on solving image classification problems. Training sets are used to train deep neural networks, and deep neural networks can extract the features of the data set pictures themselves and continue to adjust model parameters until the training is completed. This paper uses Google's open source framework TensorFlow as the experimental deep learning platform to conduct intelligent classification research on multiple annotation datasets of multiple Atlases. Manually obtain annotation images from the Atlas as sample datasets to construct a convolutional neural network model and try to use two methods of mixed training and separate training to train the models. Experiments show that the model obtained by the mixed training method performs better.