As an important generation model, variational autoencoder plays an important role in image feature extraction, text generation, and text compression. In this paper, from the perspective of feature expression, we mainly study the representation ability and stability of variational autoencoder for image features. We extract the features from the original pixels and the normalized pixels of the image respectively. Through the performance of the image classification task, we evaluate the representation ability of the variational autoencoder and compared with the traditional methods of dimensionality reduction — principal components analysis, autoencoder. The experiments on multiple datasets prove that variational autoencoder is a new non-linear dimensionality reduction method, which can represent the data effectively and stably.