Φ-GAN: Physics-Inspired GAN for Generating SAR Images Under Limited Data

Conference: ICCV 2025 arXiv: 2503.02242 Code: N/A Area: Other Keywords: SAR image generation, GAN regularization, point scattering center model, physical constraints, few-shot learning

TL;DR

This paper proposes Φ-GAN, which integrates the ideal Point Scattering Center (PSC) electromagnetic scattering physical model into GAN training as a differentiable neural module. Through a dual physics loss (generator physical consistency constraint + discriminator electromagnetic feature distillation), the method significantly improves the quality and stability of SAR image generation under data-scarce conditions.

Background & Motivation

Synthetic Aperture Radar (SAR) is critical in remote sensing due to its all-weather, all-day imaging capability. However, SAR images exhibit unique electromagnetic scattering characteristics and are costly to annotate, making large-scale datasets scarce. This motivates the use of GANs for SAR image generation, which faces three major challenges:

Extreme data scarcity: In practice, each class may contain only dozens of images (e.g., 5% of MSTAR yields only 121 images; 1% of OpenSARShip yields only 46 images), far fewer than natural image datasets.

Failure of conventional augmentation: The electromagnetic scattering characteristics of SAR targets vary significantly with azimuth angle ("target rotation" ≠ "image rotation"), making standard augmentations such as rotation ineffective or even harmful for SAR images.

Lack of physical consistency: Existing data-driven generative models have no knowledge of SAR imaging physics, and the generated images may visually resemble SAR imagery while being physically inconsistent.

These challenges are empirically confirmed: DiffAugment degrades significantly on SAR (FID from 290 to 1089), and ADA provides almost no improvement. The core insight is that the electromagnetic scattering model (PSC) for SAR images provides domain prior knowledge unavailable to natural image augmentation methods, and integrating it into GAN training can simultaneously constrain the generator to produce physically consistent images and prevent the discriminator from overfitting to speckle noise.

Method

Overall Architecture

Φ-GAN extends a standard conditional GAN (cGAN) with three new components: (1) a physics-inspired neural module \(\mathcal{F}_{\text{est}}\) that estimates PSC physical parameters; (2) a PSC physical model \(\mathcal{F}_{\text{phy}}\) that reconstructs electromagnetic scattering features from physical parameters; and (3) a dual-discriminator structure (\(\mathcal{D}_{\text{img}}\) evaluating raw images + \(\mathcal{D}_{\text{phy}}\) evaluating physically reconstructed results). The outputs of both the generator and discriminator jointly consider judgments in the image domain and the physical domain:

\[d_{out}(u, s) = \alpha \mathcal{D}_{\text{img}}(u) + (1-\alpha) \mathcal{D}_{\text{phy}}(s)\]

Conditional inputs include target class (one-hot encoding) and azimuth angle (encoded via Cyclic High-frequency Embedding, CHE):

\[r(\theta) = [\sin\theta, \cos\theta, \sin 2\theta, ..., \sin 5\theta, \cos 5\theta]\]

Key Designs

1. Ideal Point Scattering Center (PSC) Model: At high frequencies, a SAR target can be approximated as the superposition of \(N\) independent point scatterers, where each PSC response is a function of radar frequency \(f\) and azimuth angle \(\phi\):

\[E(f, \phi) = \sum_{i=1}^{N} A_i \cdot \exp\left(-j\frac{4\pi f}{c}(x_i \cos\phi + y_i \sin\phi)\right)\]

where \(A_i\) is the scattering amplitude and \((x_i, y_i)\) determines the PSC location. This model carries explicit physical meaning: \(A_i\) reflects the scattering characteristics of the target geometry, and the positional parameters reflect the spatial distribution of the target.

1. Physics-Inspired Neural Module \(\mathcal{F}_{\text{est}}\): Efficiently estimating PSC parameters from SAR images is the key bottleneck for integrating the physical model into GAN training. The paper formulates PSC parameter estimation as a sparse reconstruction problem:

\[\hat{\mathbf{o}} = \arg\min_{\mathbf{o}} \|\mathbf{\Psi}\mathbf{o} - \mathbf{r}\|_2 + \lambda\|\mathbf{o}\|_1\]

where \(\mathbf{\Psi}\) is a dictionary matrix encoding positional information, \(\mathbf{r}\) is the input image, and \(\mathbf{o}\) is the sparse PSC response vector. The Half-Quadratic Splitting (HQS) algorithm is unrolled into a two-stage neural network, with each stage admitting a closed-form solution:

\[\mathbf{o}^{(k)} = \mathbf{p}^{(k-1)} + \mu^{(k)} \mathbf{\Psi}^H (\mathbf{\Psi}\mathbf{\Psi}^H)^{-1} (\mathbf{r} - \mathbf{\Psi}\mathbf{p}^{(k-1)})\]

\[\mathbf{p}^{(k)} = S_{\rho^{(k)}}(\mathbf{p}^{(k-1)} + t^{(k)} \mathbf{\Psi}^H (\mathbf{\Psi}\mathbf{p}^{(k-1)} - \mathbf{o}^{(k)}))\]

where \(S_\rho\) denotes the soft-thresholding function. The fixed parameters \(t, \rho, \mu\) of the conventional HQS algorithm are replaced by learnable parameters at each stage, resulting in only 6 trainable parameters in total across both stages, enabling effective optimization even under extreme data scarcity.

1. Dual Physics Loss:

Generator physics loss \(\mathcal{L}_{\text{phy}}^G\): Constrains the generated images to have physical parameters consistent with real images, including both image-level and feature-level terms: \(\mathcal{L}_{\text{phy}}^G = \beta \cdot \text{MSE}(s, \tilde{s}) + \gamma \sum_{i=1}^M \frac{1}{C^i H^i W^i} \|F_{\text{phy}}^i(s) - F_{\text{img}}^i(\tilde{u})\|_2\)

Discriminator physics loss \(\mathcal{L}_{\text{phy}}^D\): By distilling the electromagnetic features of \(\mathcal{D}_{\text{phy}}\) into \(\mathcal{D}_{\text{img}}\), this loss forces the image discriminator to make decisions based on electromagnetic scattering features rather than speckle noise patterns: \(\mathcal{L}_{\text{phy}}^D = \gamma \sum_{i=1}^M \frac{1}{C^i H^i W^i} (\|F_{\text{img}}^i(\tilde{u}) - F_{\text{phy}}^i(\tilde{s})\|_2 + \|F_{\text{img}}^i(u) - F_{\text{phy}}^i(s)\|_2)\)

Loss & Training

• Total loss: \(\mathcal{L}^G = \mathcal{L}_{\text{ori}}^G + \mathcal{L}_{\text{phy}}^G\), \(\mathcal{L}^D = \mathcal{L}_{\text{ori}}^D + \mathcal{L}_{\text{phy}}^D\)
• Hyperparameters: \(\alpha=0.6\), \(\beta=1\), \(\gamma=10\)
• Optimizer: Adam with learning rate 0.0002; \(\mathcal{F}_{\text{est}}\) uses AdamW with learning rate 0.002
• Training strategy: In the early stage, the discriminator is trained 5 times for every 1 generator update
• \(\mathcal{F}_{\text{est}}\) is pretrained independently before GAN training, after which its parameters are frozen
• PSC module loss: \(\mathcal{L}_{\text{PSC}} = \|\mathbf{r} - \mathbf{\Psi}\mathbf{o}^{(K)}\|_2^2 + \lambda_1\|\mathbf{o}^{(K)} - \mathbf{p}^{(K)}\|_2^2 + \lambda_2\|\mathbf{p}^{(K)}\|_1\)

Key Experimental Results

Main Results

Comparison based on ACGAN on 10% MSTAR dataset (237 images, 10 classes):

Method	SSIM↑	VIF↑	FSIM↑	GMSD↓	FID↓	KID↓
ACGAN	0.3224	0.0386	0.7432	0.1510	290.05	0.4548
+ADA	0.2606	0.0243	0.7171	0.1643	320.59	0.4240
+DA(DiffAug)	0.2168	0.0188	0.6570	0.2018	1089.47	1.8056
+DIG	0.3279	0.0311	0.7283	0.1570	373.07	0.6243
+Ours	0.3583	0.0781	0.7622	0.1385	87.27	0.0414

Ablation Study

Incrementally adding each loss component on 10% MSTAR:

Configuration	SSIM↑	FSIM↑	GMSD↓	FID↓	KID↓
ACGAN (baseline)	0.3224	0.7432	0.1510	290.05	0.4548
+\(\mathcal{L}_{\text{phy/s}}^G\)	0.3514	0.7611	0.1392	97.45	0.0671
+\(\mathcal{L}_{\text{phy/f}}^G\)	0.3477	0.7525	0.1414	99.93	0.0568
+\(\mathcal{L}_{\text{phy}}^D\)	0.3467	0.7526	0.1422	92.76	0.0471
Ours (Full)	0.3583	0.7622	0.1385	87.27	0.0414

Generalization across backbone models (StyleGAN as baseline):

Baseline + Method	10% MSTAR FID↓	5% MSTAR FID↓	1% OpenSARShip FID↓
StyleGAN	290.64	339.28	45.10
StyleGAN+Ours	174.83	305.05	44.78

Key Findings

• DiffAugment severely degrades performance on SAR (FID from 290 to 1089), validating the argument that natural image methods are not suitable for SAR
• Φ-GAN reduces FID from 290 to 87.27 (approximately 70% reduction) and KID from 0.4548 to 0.0414 (approximately 91% reduction)
• Each physics loss component is independently effective: \(\mathcal{L}_{\text{phy/s}}^G\), \(\mathcal{L}_{\text{phy/f}}^G\), and \(\mathcal{L}_{\text{phy}}^D\) each reduce FID to below approximately 100
• The physics-inspired neural module has only 6 trainable parameters, incurring negligible computational overhead
• The method consistently improves performance across three baseline architectures (CVAE-GAN, ACGAN, StyleGAN), demonstrating its generality
• The method performs well under extreme few-shot conditions such as 1% OpenSARShip (46 images) and 14% SAR-Airplane (20 images)

Highlights & Insights

• Exemplary fusion of physical modeling and deep learning: Rather than using the physical model as pre- or post-processing, the paper transforms the physical solving process into a differentiable neural module via algorithm unrolling, enabling end-to-end training.
• Symmetric elegance of the dual-loss design: The generator-side loss enforces physical consistency, while the discriminator-side loss prevents overfitting to speckle noise; the two are complementary.
• The physical module with only 6 trainable parameters embodies a "less is more" design philosophy — under extreme data scarcity, introducing strong priors is more effective than increasing parameter count.
• The PSC model is a well-established physical tool in the SAR domain; its integration approach carries methodological generality — any domain with a mature physical model can adopt a similar strategy to incorporate it into generative models.

Limitations & Future Work

• The PSC model only describes the idealized case of point scatterers, with limited ability to model distributed scatterers such as ground clutter.
• The current work is validated only on conditional GANs; extension to diffusion models is a natural direction for future work.
• The physical module requires pre-constructing the dictionary matrix \(\mathbf{\Psi}\), which must be reconfigured for different SAR system parameters.
• The method is primarily evaluated on military target recognition scenarios (e.g., MSTAR); its effectiveness in civilian SAR scenes (e.g., ships, buildings) warrants further investigation.

Related Work & Insights

• The proposed method is orthogonal to data augmentation (ADA, DiffAugment) and regularization (DIG, RLC) approaches and can be combined with them.
• CAE and CVAE-GAN focus on SAR-specific generative architecture design, while Φ-GAN focuses on physical constraints; the two directions are complementary.
• The algorithm unrolling idea originates from signal processing works such as LISTA; this paper represents a novel application of that idea in the context of generative models.
• The work provides a successful example of Physics-Informed AI in the GAN domain.

Rating

• Novelty: ⭐⭐⭐⭐⭐ First end-to-end integration of an electromagnetic scattering physical model into GAN training, with an elegant physical module design.
• Experimental Thoroughness: ⭐⭐⭐⭐ Three datasets, two backbone architectures, comprehensive ablation, and comparison with SAR-specific methods.
• Writing Quality: ⭐⭐⭐⭐ Physical background is clearly introduced, and the methodological derivations are rigorous.
• Value: ⭐⭐⭐⭐ Directly valuable to the SAR image generation community; the physical integration paradigm is transferable to other domains.

Related Papers

• [ICCV 2025] Auto-Regressively Generating Multi-View Consistent Images (MV-AR)
• [ICCV 2025] Is Meta-Learning Out? Rethinking Unsupervised Few-Shot Classification with Limited Entropy
• [ICCV 2025] Hi3DGen: High-fidelity 3D Geometry Generation from Images via Normal Bridging
• [ICCV 2025] Recover Biological Structure from Sparse-View Diffraction Images with Neural Volumetric Prior
• [ICCV 2025] NAPPure: Adversarial Purification for Robust Image Classification under Non-Additive Perturbations