SAR Vehicle Data Generation With Scattering Features for Target Recognition

As is well known, obtaining high-quality measured SAR vehicle data is difficult. As a result, deep learning-based data generation is frequently utilized for SAR target augmentation because of its affordability and simplicity of use. However, existing methods do not adequately consider the target scattering information during data generation, resulting in generated target SAR data that does not conform to the physical scattering laws of SAR imaging. In this paper, we propose a SAR target data generation method based on target scattering features and Cycle-Consistent Generative Adversarial Networks (CycleGAN). First, a physical model-based method called Orthogonal Matching Pursuit (OMP) is adopted to extract the Attribute Scattering Centers (ASC) of SAR vehicle targets. Then, a multidimensional SAR target feature representation is constructed. Based on the scattering difference between the generated and real SAR target images, we introduce a loss function and further develop a generative model based on the CycleGAN. Therefore, the scattering mechanisms of SAR targets can be well learned, making the generated SAR data conform to the target scattering features. We conduct SAR target generation experiments under standard operating conditions (SOC) and extended operating conditions (EOC) on our self-acquired dataset as well as SAMPLE and MSTAR datasets. The SAR vehicle target data generated under SOC shows a more accurate scattering feature distribution to the real target data than other state-of-the-art methods. In addition, we generate SAR target data under EOC that conforms to SAR imaging patterns by modulating ASC feature parameters. Finally, the target recognition performance based on our proposed generated SAR vehicle data under SOC is validated, where the recognition rate increased by 4% after the addition of our generated target data. The code for the proposed method is publicly available at https://github.com/freel3/ Transformation _Mstar2SAMPLE.