DATA-DRIVEN SAR IMAGE RECONSTRUCTION METHOD USING ADMM WITH DEEP PLUG-AND-PLAY PRIORS

Synthetic Aperture Radar (SAR) image reconstruction fundamentally constitutes an ill-posed reconstruction conundrum, with the overarching objective of faithfully restoring the'authentic' SAR scenes via radar echo measurements and a forward model. The conventional FFT-based approach, in conjunction with matched filtering, tends to yield undesirable sidelobes and speckle artifacts. Conversely, the reconstruction challenge can be cast as a minimization problem, seeking to enhance outcomes by merging data-term and prior-related terms. This composite problem is amenable to resolution via proximal methods. In this paper, a novel data-driven methodology is employed, leveraging the plug-and-play (PNP) technique to train a neural network for acquiring target image priors. In comparison to traditional manually crafted priors and PNP priors, the proposed PNP deep priors exhibit the capacity to achieve superior reconstruction results. Experimental assessments involving synthetic and real SAR scenes underscore the robustness and efficacy of the proposed approach.