Synthetic Aperture Radar Imaging Meets Deep Unfolded Learning: A comprehensive review

Synthetic aperture radar (SAR) can obtain high-resolution images without being affected by environmental visibility. The compressed sensing (CS) technique is considered to be a strong candidate for simplifying SAR system complexity and improving imaging quality. With CS, SAR imaging is addressed by optimizing an augmented object function with data fidelity and feature-oriented priors; however, it suffers from time-consuming calculations and poor adaptability.

Recently, an emerging technique dubbed deep unfolded/unrolled learning, or model-driven learning, offers promise in eliminating such issues by bridging the gap between the learning framework and iterative algorithms. The increasing popularity of unfolded networks in SAR inverse problems also shows their potential for developing efficient and accurate imaging algorithms. This article surveys the SAR imaging algorithms based on deep unfolding techniques. We extensively cover different imaging regimens including conventional 2D SAR, inverse SAR (ISAR), 3D SAR, and automotive radar imaging. On the algorithm side, deep unfolding frameworks are mainly categorized according to the feature-oriented regularizers, and their characteristics, principles, and feasibility in SAR inverse problems are discussed in detail. By reviewing pioneering works, we discuss and reveal the current research stages in different tasks. Finally, the limitations, challenges, and opportunities of deep unfolding techniques are discussed in different radar imaging tasks.