$\mathbfΦ$-GAN: Physics-Inspired GAN for Generating SAR Images Under Limited Data

TL;DR

Φ-GAN addresses SAR image generation under limited data by embedding the electromagnetic PSC model into GANs through a physics-inspired neural module and two dedicated physical losses. The framework uses a dual-discriminator setup to enforce both image fidelity and physics-consistent EM features, enabling end-to-end learning with few samples. Across multiple SAR datasets, Φ-GAN demonstrates improved stability, faster convergence, and superior generalization compared to baseline GANs and SAR-specific methods, with ablations confirming the contribution of each physical regularizer and the PSC estimator. This physics-aware regularization enhances the realism and physical plausibility of generated SAR images, offering practical benefits for SAR analysis and downstream tasks when data are scarce.

Abstract

Approaches for improving generative adversarial networks (GANs) training under a few samples have been explored for natural images. However, these methods have limited effectiveness for synthetic aperture radar (SAR) images, as they do not account for the unique electromagnetic scattering properties of SAR. To remedy this, we propose a physics-inspired regularization method dubbed $Φ$-GAN, which incorporates the ideal point scattering center (PSC) model of SAR with two physical consistency losses. The PSC model approximates SAR targets using physical parameters, ensuring that $Φ$-GAN generates SAR images consistent with real physical properties while preventing discriminator overfitting by focusing on PSC-based decision cues. To embed the PSC model into GANs for end-to-end training, we introduce a physics-inspired neural module capable of estimating the physical parameters of SAR targets efficiently. This module retains the interpretability of the physical model and can be trained with limited data. We propose two physical loss functions: one for the generator, guiding it to produce SAR images with physical parameters consistent with real ones, and one for the discriminator, enhancing its robustness by basing decisions on PSC attributes. We evaluate $Φ$-GAN across several conditional GAN (cGAN) models, demonstrating state-of-the-art performance in data-scarce scenarios on three SAR image datasets.

Figures (6)

• Figure 1: (a) Our goal is to generate SAR images at unseen angles using a few samples. (b) The unique electromagnetic imaging mechanism of SAR creates a significant discrepancy between “image rotation” and “target rotation”. (c) and (d) show SAR images generated by DIGGAN fang2022diggan, while (e) and (f) display images generated by our proposed $\Phi$-GAN. It is evident that our method produces well-detailed results both near seen angles (red) and at unseen angles (blue). All SAR images are single-channel and visualized using the “hot” colormap for enhanced visual interpretation.
• Figure 2: (a) The proposed $\Phi$-GAN framework overview. (b) The demonstration of the physics-inspired neural module ($\mathcal{F}_{\mathrm{est}}$) for physical parameter inversion and image reconstruction based on the point scattering center (PSC) model ($\mathcal{F}_{\mathrm{phy}}$). $\mathcal{F}_{\mathrm{est}}$ contains a couple of learnable parameters and $\mathcal{F}_{\mathrm{phy}}$ is an off-the-shelf reconstruction process. (c) The illustration of PSC model $\mathcal{F}_{\mathrm{phy}}$. (d) The implementation of the $k$-th stage within $\mathcal{F}_{\mathrm{est}}$.
• Figure 3: The discriminator output of ACGAN with and without our proposed physics-inspired regularization.
• Figure 4: Results on OpenSARShip and SAR-Airplane.
• Figure 4: The generated SAR images and their associated PSC model reconstruction results during model training (from 200$^{th}$ epoch to 2000$^{th}$ epoch). (a) and (c) are the generated SAR images of ACGAN and $\Phi$-ACGAN, respectively. (b) and (d) are the PSC reconstruction results of (a) and (c), respectively. (e) and (f) are the training image and its PSC reconstruction result.