SG2D-AFnet: Structure-Guided 2-D Autofocus Network for Microwave Photon SAR

Microwave Photon Synthetic Aperture Radar (SAR) enables large-scene remote sensing observations with centimeter-to even millimeter-level resolution. However, its ultra-high resolution characteristic poses notable bottlenecks for existing autofocus methods. Traditional autofocus methods based on physical or statistical models suffer from insufficient robustness or low efficiency. On the other hand, existing Deep Neural Network (DNN)-based autofocus methods neglect the imaging physical model due to their pure data-driven nature, resulting in suboptimal autofocus performance. To address these issues, this paper proposes a structure-guided two-dimensional (2-D) autofocus network tailored for ultra-high resolution microwave photon SAR. Its core breakthroughs are as follows: i) During the deep network training process, an innovative coupling relationship is established among the range defocus parameters, residual Range Cell Migration (RCM) parameters, and azimuth defocus parameters in the defocused image. By leveraging the analytical mapping relationship among these three, the 2-D (range-azimuth) parameter estimation is simplified to a one-dimensional (1-D) problem that only requires estimating the azimuth defocus parameters, significantly accelerating the stable convergence of the autofocus network. ii) It specifically solves the range defocus problem that existing 2-D autofocus networks commonly fail to consider, effectively improving the range resolution. The processing results of measured data from a Ku-band airborne microwave photon SAR with a resolution of 0.03 meters show that the proposed autofocus network achieves excellent focusing effects in two typical scenarios, i.e., the prominent point target scenario and the uniformly distributed target scenario. It effectively overcomes the limitations of traditional autofocus methods and existing DNN-based autofocus methods.