Traditional network traffic classification methods are difficult to distinguish network traffic encrypted using VPN. In order to achieve classification of encrypted traffic, a network traffic image classification method based on multiple composition methods is proposed. Five special composition methods are studied to convert encrypted network traffic into traffic images, and finally convolutional neural networks are used to classify them. The experimental results on self-collected VPN encrypted traffic dataset and ISCX VPN-nonVPN public dataset show that the classification accuracy of this encrypted traffic classification method reaches more than 90% and 95% on the two datasets, respectively. The classification accuracy of diagonal or waterfall composition method has about 1% improvement over the traditional line composition method. The special composition method achieves the improvement of encrypted traffic classification accuracy by enhancing the correlation of pixel points in the traffic image and increasing the image features in the traffic image.