SAR image generation via statistical regularization and scattering feature guidance

作者：Yun Liang, W W Wang, R H Jin, J LIU, Bingwei Hui · 发表于：IET conference proceedings. · 年份：2026 · DOI：10.1049/icp.2026.1468 · 研究领域：Advanced SAR Imaging Techniques、Synthetic Aperture Radar (SAR) Applications and Techniques、Advanced Neural Network Applications

Generative models for Synthetic Aperture Radar (SAR) imagery often suffer from a mismatch with the statistical characteristics of real data and a lack of physical interpretability. To address these limitations, we propose a novel framework for vehicle target generation that integrates both rigorous statistical regularization and physics-informed scattering constraints. Our approach is based on a Generative Adversarial Network (GAN). Crucially, we introduce a Yeo-Johnson transform module to adaptively normalize the non-Gaussian SAR data into an approximately Gaussian distribution prior to network training. This transformation enhances model stability and ensures the statistical fidelity of the generated imagery, which is restored to the original data domain via an inverse transform. To enforce physical realism, the generator leverages a capsule network to preserve target structural integrity. Furthermore, we introduce a scattering feature extraction module and a corresponding loss function that constrains the generative process to adhere to electromagnetic scattering laws. Experiments on the public MSTAR and a newly-collected dataset demonstrate that images generated by the proposed method are highly consistent with real targets in both their statistical and scattering properties. Downstream target recognition experiments validate that augmenting a training set with our generated data effectively improves the recognition rate by 5.91%, confirming the practical utility of our a...