Complex-valued SAR image super-resolution via frequency-domain cold diffusion

Abstract

Synthetic Aperture Radar (SAR) can acquire high-quality data under all-time and all-weather conditions, and thus possesses important application value in multiple fields such as national defense, remote sensing and disaster monitoring. Nevertheless, the imaging resolution of SAR is constrained by the bandwidth of transmitted signals and the length of the synthetic aperture. Hardware modifications for obtaining high-resolution (HR) SAR images will lead to extremely high costs. To address this problem, researchers have adopted SAR super-resolution (SR) technology to improve the resolution of SAR images. However, relevant research in this field is still limited, because it is extremely difficult to obtain accurately registered pairs of HR and low-resolution (LR) SAR images. Existing LR image generation approaches also cannot characterize the essential nature of SAR image resolution degradation. Furthermore, SR based on general images are not well adapted to SAR-specific features, thereby constraining model generalization and the credibility of output results. To solve the above problems, we propose a feasible method. The method is based on cold diffusion: in the forward process, a high-frequency component mask is used as a degradation operator to construct a deterministic degradation chain in the two-dimensional frequency domain that aligns with the SAR image resolution degradation mechanism; in the reverse process, a parametric recovery network is designed and implemented, and SR reconstruction of LR observations is achieved step-by-step through an iterative inference strategy. A series of experimental results demonstrate that the proposed model outperforms existing SR methods in both quantitative and qualitative evaluations.