Convolutional neural networks (CNNs) in deep learning enable the extraction of features in image data. Through the multi-layer superposition of a convolutional neural network, we can better capture the essential characteristics of different auroral subclasses and further classify auroral images in detail. Because the auroral morphological features often present abstract characteristics, our study compares different CNN architectures and different layering in order to test the best neural network model for mesoscale aurora classification. Although the classification models and subclasses used by us are both more complex, the highest F1 score of aurora classification of the test set reaches 99.6% (ResNet-50), which performs best comparing with previous works. Our classification models work also quite well when applied to an independent auroral image sequence, declaring our approach can automatically select images of various mesoscale auroral forms using CNNs, and allow the time sequence of auroral evolution to be seen automatically through the mesoscale auroral feature recognitions.