Improved SAR Image Generation with Double Top-K Training Method on Auxiliary Classifier GAN

Synthetic aperture radar (SAR) is a critical imaging technique that is widely used for civil and military tasks, as it is featured with an excellent ability for high resolution imaging. However, due to the severe shortage of SAR images, the performance of automatic target recognition (ATR) is greatly sabotaged. Generative adversarial network (GAN) is often applied for data augmentation of small-sized dataset. In this paper, based on auxiliary classifier GAN (ACGAN) and top-k training technique, we propose double top-k training, which implements a modification during training without any further adjustment on model architecture. The proposed method is to enforce generator to only optimize on generated images that perform well in both discriminator and auxiliary classifier, and discard images of poor performance. We evaluate the generated images via recognition on the moving and stationary target acquisition and recognition (MSTAR) dataset. Recognition accuracy and Fréchet inception distance (FID) score indicate better generation results of the proposed method compared with original ACGAN.