In order to solve the problem that the recognition performance is obviously degraded when the model trained by known data distribution transfer to unknown data distribution, domain generalization method based on attention mechanism and adversarial training is proposed. Firstly, a multi-level attention mechanism module is designed to capture the underlying abstract information features of the image; Secondly, increases the loss limit of the generative adversarial network，the virtual enhanced domain which can simulate the target domain of unknown data distribution is generated by adversarial training on the premise of ensuring the consistency of data features and semantics; Finally, through the data mixing algorithm, the source domain and virtual enhanced domain are mixed and input into the model to improve the performance of the classifier. The experiment is carried out on five classic digit recognition and CIFAR-10 series datasets. The experimental results show that the model can learn better decision boundary, generate virtual enhanced domain and significantly improve the accuracy of recognition after model transplantation. Comparing to the previous method, our method improves average accuracy by at least 2.5% and 3% respectively. Experiments on five classic digit recognition and CIFAR-10 series datasets which significantly improves the classification average accuracy after model transfer.