BS-ACGAN: A High-Resolution Multi-Class SAR Image Generation Approach via Conditional GAN

Deep learning-based synthetic aperture radar (SAR) target recognition and detection technologies have demonstrated significant potential. However, challenged by the unique imaging mechanism and complex scattering characteristics of SAR systems, acquiring large-scale high-quality SAR image datasets poses substantial challenges, which severely limits model training and performance improvements, which in turn reduces detection effectiveness. To address these issues, this paper proposes a self-attention-enhanced high-resolution multi-category SAR image generative adversarial network named Big Self-attention Auxiliary Classifier Generative Adversarial Network (BS-ACGAN). The model builds upon the BigGAN architecture by embedding self-attention modules into multi-scale feature layers of both generator and discriminator, strengthening cross-category global feature relationship modeling. Simultaneously, it incorporates the conditional adversarial training mechanism of the ACGAN framework to achieve multi-category generation. Experimental results on both a self-built ship dataset and the e Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset demonstrate that compared with Re-ACGAN and ACGAN, BSACGAN delivers superior generation performance. In classification tasks using expanded datasets, the average accuracy shows significant improvement over the original dataset and outperforms existing comparative models.