An improved cycle-consistent adversarial networks （CycleGAN）-based method for augmenting data of maize leaf disease images was designed to solve the difficulties in obtaining image dataset of recognizing maize diseases，insufficient samples，and imbalanced samples across different categories of diseases.Convolutional kernels with smaller receptive fields were used to optimize the structure of the CycleGAN network and generate sample images with high-quality and reduce the occurrence of overfitting.The SE （squeeze-excitation） attention mechanism was embedded into the residual module of the generator to enhance the ability of CycleGAN to extract disease features and allow the network to more accurately capture diseases with small target or features with subtle inter-domain differences.The results showed that the improved CycleGAN reduced the frechet inception distance （FID） value of generated disease images by 43.33，32.67，24.24，and 19.72 compared with the original CycleGAN，DCGAN，DCGAN+，and WGAN algorithms. GAN-train and GAN-test increased by 3.13 and 4.25 percentage points compared to that in the original CycleGAN.The improved CycleGAN-based method for augmenting data was used to construct a dataset of maize diseases，and the accuracy of recognizing the maize leaf diseases based on this dataset was significantly improved on three network architectures.AlexNet，VGGNet，and ResNet increased by 3.90，4.41，and 3.44 percentage points，respectively.Compared with traditional data augmentation algorithms，the disease recognition rate of ResNet network architecture increased by 5.79 percentage points.It is indicated that the improved CycleGAN network effectively solves the problem of insufficient image dataset of maize diseases.