ISAGAN: A High-Fidelity Full-Azimuth SAR Image Generation Method

Abstract

The amount of data is often insufficient in the research of intelligent interpretation of synthetic aperture radar (SAR) images. Recent researches show that generative adversarial network (GAN) has achieved excellent results in data augmentation, so it is commonly used to solve the few samples problem. However, GAN is mainly used to generate optical images instead of SAR images, and its training process is unstable. To address these issues, the improved self-attention GAN (ISAGAN) is proposed to generate high-fidelity full-azimuth SAR images. First, simplified self-attention GAN is proposed for high-quality SAR image generation. Second, the azimuth and class of the SAR image are added to the network, enabling the network to generate SAR images with arbitrary azimuth and class. Last, instance noise is added to the training process, which improves the training stability of the network. The MSTAR dataset is used to verify the effectiveness of the proposed SAR image generation method. The experiment results confirm that the ISAGAN can generate high-quality multi-class full-azimuth SAR images.