Frequency-enhanced diffusion prior network for joint SAR image despeckling and super-resolution

Abstract

Synthetic aperture radar (SAR) image enhancement faces inherent difficulties in simultaneous speckle suppression and structural detail preservation, severely limiting the performance of conventional methods in high-precision remote sensing tasks. This paper proposes a novel two-stage cascaded deep learning framework for high-quality SAR image reconstruction, which sequentially integrates a window-based Transformer denoising subnet and a diffusion-driven super-resolution subnet via a learnable projection matrix. To address the coupled degradation of speckle noise, imaging blurring, and structural loss, four dedicated designs are adopted, including a projection-coupled serial architecture for joint optimization and lightweight parameter deployment (24.3% parameter reduction), a Frequency Enhancement Module for high-frequency detail recovery, an Adaptive Noise Scheduler for robust diffusion adjustment under complex textures and low-contrast conditions, and Meta Residual Connections for stable deep feature propagation. Quantitative experiments on the FAIR-CSAR-V1.0 ×4 super-resolution benchmark demonstrate that the proposed method achieves state-of-the-art performance with 19.26 dB PSNR, 0.3502 SSIM, 2.23 ENL, and 2.26 RadRes. Our method outperforms the baseline model by 5.71% in PSNR and 30.4% in SSIM, and surpasses existing SOTA methods by 31.9%, 108.2%, 105.6%, and 27.3% in four metrics, respectively, with prominent gains mainly obtained in challenging noisy and low-contrast SAR regions. Ablation studies validate the efficacy of each component. The proposed framework provides an effective lightweight solution for high-precision SAR reconstruction applicable to military reconnaissance, geological exploration, and disaster monitoring.