Answers are grounded in this paper and cite their sources.
Current synthetic aperture radar (SAR) image generation methods for data augmentation suffer from three fundamental limitations: insufficient continuous aspect-angle control precision, ineffective heterogeneous condition fusion, and compromised complex-valued physical fidelity. These limitations hinder effective SAR automatic target recognition (ATR) data augmentation. To address these challenges, this article proposes a Physics-Constrained Controllable Diffusion Model for SAR-ATR Data Augmentation (PC-DiffSAR). First, a time-step-modulated aspect encoding mechanism is introduced. It enables precise controllable generation of continuous aspect angles through dynamic modulation adapted to different stages of the diffusion process. Subsequently, a parallel multisource cross-attention mechanism is designed. It decouples parameter optimization of heterogeneous conditions through independent key-value projection pathways, achieving joint control of target category and aspect angle. Finally, an amplitude-weighted complex-valued constraint loss is formulated, which utilizes the closed-form solution of the diffusion forward process to preserve the intrinsic coherency between real and imaginary components. This not only maintains complex-valued physical fidelity but also retains the computational efficiency of real-valued networks. Experiments on the Moving and Stationary Target Acquisition and Recognition dataset demonstrate that PC-DiffSAR outperforms existing methods in generation quality, physical fidelity, and conditional control accuracy. Data augmentation experiments confirm that under few-shot scenarios, this method improves classification accuracy by approximately 9–17 percentage points. This demonstrates the practical value of the method for few-shot SAR-ATR applications.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: