Electromagnetic Mechanisms Guided Structure-Constrained Flow Matching for Rapid Realistic SAR Image Generation

Abstract

The scarcity of real synthetic aperture radar (SAR) images, coupled with significant discrepancies between the simulated and real samples, constrains the performance of automatic target recognition (ATR). Thus, this letter proposes an efficient, physically rigorous, and lightweight realistic SAR image generation method for limited real-sample scenarios by integrating the scattering center model with flow matching. First, a GPU-accelerated forward modeling rapidly generates physically interpretable target scattering features from computer-aided design (CAD) models within tens of milliseconds. Based on this, a structure-constrained flow matching network incorporating a global self-attention mechanism is designed. Using scattering-derived physical scattering skeleton as structural prior, the network effectively reconstructs the complex scattering distributions and nonlinear background noise in real SAR images while ensuring the accuracy of the target’s electromagnetic characteristics. Experiments on the SAMPLE dataset demonstrate that this method outperforms traditional generative adversarial networks (GANs) models on key metrics (FID, KID, and LPIPS). In few-shot scenarios classification tasks, classifiers trained on our synthesized images achieve accuracy close to those of real data, demonstrating that our method effectively bridges the domain gap between synthetic and real data, enabling direct application to downstream tasks.