Azimuth-Controllable SAR Image Generation Method Based on Self-Attention Mechanism Generative Adversarial Network

In recent years, deep learning algorithms have shown significant promise in Synthetic Aperture Radar (SAR) Automatic Target Recognition (SAR-ATR); however, their performance is heavily dependent on the quality and size of the training dataset. A major challenge is the limited sample problem, which often leads to overfitting in deep neural networks when datasets are sparse. To address this, we propose a Self-attention Conditional Generative Adversarial Network (SA-CGAN) that enables precise control over the azimuth angles of generated SAR images. In our approach, the integration of a self-attention mechanism with a conditional GAN framework allows the network to dynamically focus on key regions and capture long-range dependencies, thereby enhancing the structural integrity and scattering features of the synthesized images. Furthermore, the generated images themselves can be directly used to augment existing datasets and mitigate the issue of limited samples, fostering the development of more robust ATR models. Experimental results show that SA-CGAN outperforms traditional GAN methods by generating high-quality SAR images with minimal error in both structure and scattering characteristics. Quantitative evaluations confirm that our method significantly improves the performance of SAR-ATR models, even when used to augment limited real datasets, offering a promising solution to data scarcity in various target recognition applications.