In order to solve the problems of poor generalization ability and low recognition accuracy in the identification of cucumber leaf disease with small samples under complex field environment， the self-attention mechanism module was introduced into the activation reconstruction network AR-GAN （activation reconstruction GAN）， and the smooth L₁ regularization was used as the loss function to design and improve the activation reconstruction network IAR-GAN （improved AR-GAN） to expand the cucumber leaf disease image. By adding void convolution and deformation convolution on the basis of the Inception network， the void and deformation convolution neural network （DDCNN） was designed for cucumber leaf disease identification. The test results showed that the proposed IAR-GAN effectively alleviated the over-fitting phenomenon and enriched the diversity of generated samples. The average recognition accuracy of the proposed DDCNN for cucumber anthracnose， spot target disease and downy mildew was more than 96%， which is 9% higher than the Inrception-V3 model. The above results showed that the data expansion method and disease identification model proposed in this paper can provide new ideas for the accurate identification of crop leaf diseases with small samples in complex environments.