SEIS-Net: A 3-D SAR Enhanced Imaging Network Based on Swin Transformer

Abstract

Conventional 3-D synthetic aperture radar (SAR) sparse imaging algorithms suffer from degradation in weakly sparse scenes due to their reliance on inherent sparsity. In addition, they are constrained by high computational complexity and parametric tuning. To address these problems, we propose a novel 3-D SAR enhanced imaging network based on swin transformer dubbed SEIS-Net. The proposed algorithm consists of two cascaded stages. The first one focuses on estimating the missing measurement elements by constructing a Unet based on the swin transformer. The second stage aims to recover a high-quality image from the estimated echo matrix. The proposed imaging network is theoretically derived from fast iterative shrinkage-thresholding algorithm optimization framework, where the network weights can be learned from an end-to-end training procedure. Finally, simulations and real-measured experiments are carried out. Both visual and quantitative results demonstrate the superiority of the proposed SEIS-Net over the current state-of-the-art algorithms in reconstructing 3-D images from sparsely sampled echoes.