Overview of the application of generative adversarial networks in SAR image processing

DOI:	10.11809/bqzbgcxb2024.04.002
Keywords:	synthetic aperture radar; generative adversarial networks; SAR datasets; high fidelity image generation
Abstract:	Synthetic aperture radar (SAR) automatic target recognition (ATR) technology is a research hotspot in the field of SAR image processing, but the situation of insufficient data samples leads to the limitation of SAR ATR application research. The traditional image simulation techniques for expanding SAR datasets have complex models, large computation, and the generated images are not realistic enough. Generative Adversarial Networks GANs do not need target prior information and can generate realistic images directly from real image data, which has the advantages of low loss and end to end, so it is more suitable for high quality expansion of small sample SAR data compared with traditional methods. The article focuses on the research and application of GANs in SAR image processing, and introduces the methods for acquiring target SAR images, including traditional simulation technology and GANs technology based on deep learning. The commonly used SAR datasets for GANs training are introduced from the aspects of target images and scene images. Aiming at the application scenarios of different datasets, the latest research progress of GAN networks in target SAR image generation, SAR super resolution reconstruction, SAR and optical image fusion is mainly introduced. Finally, the article combines with deep learning and SAR target characteristics, we give the suggestions for the subsequent development of GANs network in SAR image applications.
Published:	2024-04-30
PDF