Azimuth-transfer based SAR image generation with few samples

Due to the azimuth sensitivity and the limited availability of annotated samples in synthetic aperture radar (SAR) images, the SAR target recognition accuracy of deep models tends to decrease. To address this issue, this paper proposes a novel azimuth-transfer SAR image generation algorithm based on Generative Adversarial Networks (GANs). The algorithm separately extracts category and azimuth features from images, enabling the transfer of azimuth information from the source domain to the target domain, thereby augmenting samples at the missing azimuth angles in the target domain. The algorithm constructs a GAN model adapted to azimuth transfer, designs a multi-dimensional loss function system, and proposes a source domain-assisted training data construction strategy. Experimental results show that the proposed algorithm achieve superior performance compared to state-of-the-art methods. When the generated images are applied to SAR target recognition tasks, the recognition accuracy is significantly enhanced.