U-ADMMNet: A Mdel-Based Deep Learning Method for Sparse Aperture ISAR Imaging

Abstract

Sparse inverse synthetic aperture radar (ISAR) imaging is a linear undetermined problem, which is generally solved by compressed sensing (CS) or sparse signal recovery (SSR) theory at present. The existing SSR methods introduce prior information to the task of sparse aperture ISAR imaging via regularization. However, the obtained images are too sparse to provide rich structural information of the target due to the simple regular term constraint. In this paper, a novel prior framework for sparse aperture ISAR imaging is proposed, which utilizes complex-valued U-Net (CV-U-Net) networks as the denoiser prior module to learn prior information. The prior model based on the CV-U-Net network is combined with the alternating direction method of multipliers (ADMM) algorithm, which we call U-ADMM. It can bridge the weakness between learning-based methods and model-based methods. Moreover, the U-ADMM is unrolled into a deep neural network, dubbed as U-ADMMNet, which solves the limitation of the same denoiser in each iteration of the past methods. Simulation and measured experimental results show that the proposed U-ADMMNet imaging method can significantly improve the image quality from limited observations while keeping more powerful robustness and adaptability than other state-of-the-art methods.