Abstract: Aiming at the problem of imbalanced sample datasets in classification tasks and very low accuracy of classification models on minority classes, this paper proposes a cascade processing method based on EnsNet and MCGAN models for background style transfer and font style transfer. The EnsNet model can be used for the task of font erasure and font extraction of complex backgrounds, MCGAN model can be used for the task of style transfer and data expansion of the extracted fonts. On the premise of ensuring that the sample diversity is satisfied, the cascade method of the two sets of models is used for the task of the cross-order expansion of the minority samples. The results show that, first of all, the optimized LeNet5-BN sample expansion model is selected for verification. On the original real datasets with severely unbalanced data distribution, the minority class recognition accuracy is less than 99.50%. On the synthetic datasets after using the data expansion model, The original minority recognition accuracy rate increased to 99.98%. Secondly, the Resnet and Mobilenet models were used to further verify the classification and recognition accuracy of the expansion of the sample. The classification accuracy of the expansion was increased from 99.88% and 99.8% to 99.96% and 99.95%, respectively. The sample expansion effect has been well verified by multiple models. Finally, the LeNet5-BN model was selected to implement ten cross-validation experiments, and the average recognition accuracy rate increased from 99.50% to 99.98%, further indicating that the sample cross-order expansion model has perfect robustness.