This paper proposes a deep unfolding framework for high-resolution Synthetic Aperture Radar (SAR) image reconstruction under non-ideal acquisition conditions (undersampling, phase/motion mismatch, and multiplicative speckle noise). The proposed method (DU-SAR) decomposes the optimization algorithm into a series of steps with two main components: (i) a differentiable SAR physics operator-based data consistency algorithm, and (ii) a speckle-aware proximal/learned denoiser to preserve edges and textures. To address defocus due to phase errors, we embed an in-the-loop joint autofocus that updates the phase map at each unrolling step. The training scheme is two-stage—pretraining on synthetic data with varying undersampling/SNR levels and self-supervised fine-tuning on real data based on measurement domain consistency—with GPU acceleration, mixed precision, and multi-resolution unrolling for efficiency. Experimental results show consistent improvements over classical, model-based, and deep baselines end-to-end: at 50% undersampling, DU-SAR achieves a PSNR of 30.9 dB and an SSIM of 0.87, and 28.9 dB/0.83 at 25%; robustness to phase errors is maintained with an SSIM of 0.71 at an RMS error of 1.00 rad. Performance-wise, an inference latency of approximately 85 ms per 512×512 patch makes the method feasible for near real-time on mid-range GPUs. These findings confirm that physics-consistent and speckle-aware deep unfolding effectively recovers high-frequency details while maintaining focus and computational efficiency.
Ariel Moura Vilas Boas (Thu,) studied this question.