Multi-Category SAR Images Generation Based on Improved Generative Adversarial Network

The generative adversarial network (GAN) provides a different way for SAR data augmentation. The traditional GAN model is mainly based on the Jensen-Shannon (JS) divergence or Wasserstein distance. The former faces mode collapse, while the latter is not suitable for multi-category image generation. In this paper, an improved model based on WGAN-GP is proposed. An encoder is used to learn the features of real samples as the input of the generator to control training to a certain extent and make the generated image quality better. In addition, a pre-trained classifier is introduced as the constraint of the generator to ensure the generated images have the correct category information. MSTAR dataset is used to verify the generation capability of the proposed model. The results show that the proposed model has the stable generation capability to provide high-quality SAR images as a supplementary training dataset, which could assist in achieving good classification accuracy.