In the process of zero-shot image classification, problems such as high acquisition cost for samples of known categories and domain drift were addressed. A zero-shot image classification model based on maximum mean difference was proposed to improve the variational auto-encoder. First, the noise factor of samples is separated by maximizing the mean difference to obtain samples closer to the unknown category. Then, the generated sample-assisted learning is used to transform the zero-shot classification problem into the supervised learning classification problem. Finally, the classification model is used for image classification. Compared with the zero-shot image classification algorithm based on distance measurement, the proposed algorithm achieved good classification effect on CUB, AWA, and ImageNet-2 data sets, and improved domain drift and classification accuracy, which proves the effectiveness and feasibility of the proposed algorithm model.