Unsupervised 3D SAR Imaging Network Based on Generative Adversary Learning

Reconstructing 3-D synthetic aperture radar (3D SAR) images from the sparsely sampled echo measurements holds significant importance in simplifying the system complexity and sensing costs. The conventional compressed sensing (CS)-based imaging algorithms address the problem by suspecting the sparsity of the imaging space, showing good reconstruction performance but failing to tackle the imaging tasks in weakly sparse scenes. Besides, traditional iterative imaging algorithms also suffer from high computational complexity, cumbersome parameter tuning, and poor adaptability. To address these issues, a novel unsupervised 3D SAR image reconstruction network is proposed for estimating 3D SAR images from the corresponding incomplete echo measurements. The proposed scheme contains two stages. Wherein, the first stage aims to estimate the missing echo elements from the incomplete observations by designing a generative adversary network based on partial convolution-based generative adversarial network (PCGAN). The second phase focuses on reconstructing the target synthetic aperture radar (SAR) image from the estimated echoes by constructing an unsupervised adaptive fast iterative shrinkage-thresholding algorithm (FISTA)-inspired deep unfolding network (AdaFIST-Net). Finally, simulations and real-measured experiments are carried out. Experimental results show that the proposed imaging network outperforms the current state-of-the-art algorithms in reconstructing 3-D images from sparsely sampled echoes in various sampling conditions and SNR cases.