一种基于元学习的稀疏孔径ISAR成像算法  

A Metalearning-based Sparse Aperture ISAR Imaging Method

作　　者：夏靖远 杨志雄 周治兴 廖淮璋 张双辉 付耀文[1] XIA Jingyuan;YANG Zhixiong;ZHOU Zhixing;LIAO Huaizhang;ZHANG Shuanghui;FU Yaowen(National University of Defense Technology,Changsha 410073,China)

机构地区：[1]国防科学技术大学,长沙410073

出　　处：《雷达学报（中英文）》2023年第4期849-859,共11页Journal of Radars

基　　金：国家自然科学基金(62171448,61921001,62131020,62022091),湖南省杰出青年基金(2022JJ10067)。

摘　　要：稀疏孔径逆合成孔径雷达(ISAR)成像的目标是从不完整的回波中恢复和重建高质量ISAR图像,现有方法主要可以分为基于模型的方法和基于深度学习的方法两大类:一方面,基于模型的稀疏孔径ISAR成像方法往往具备显性的数学模型,对雷达回波的成像过程有清晰的物理建模,但算法有效性上不如基于学习的方法。另一方面,基于深度学习的方法通常高度依赖训练数据,难以适配空间目标ISAR成像任务中高实时、高动态的现实应用需求。针对上述问题,该文提出了一种基于元学习的高效、自适应稀疏孔径ISAR成像算法。所提方法主要包含基于学习辅助的交替迭代优化和元学习优化两部分。基于学习辅助的交替迭代优化继承了ISAR成像机理的回波成像模型,保证了方法数学物理可解释性的同时避免了方法对数据的依赖性;基于元学习的优化策略通过引入非贪婪优化策略,提高了算法跳出局部最优解的能力,保证了病态非凸条件下的算法收敛性能。最后,实验结果表明:该文方法可以在不依赖训练数据、不进行预训练的情况下实现高效、自适应的稀疏孔径ISAR成像,并取得优于其他常规ISAR成像算法的性能。Sparse Aperture-Inverse Synthetic Aperture Radar(SA-ISAR)imaging methods aim to reconstruct high-quality ISAR images from the corresponding incomplete ISAR echoes.The existing SA-ISAR imaging methods can be roughly divided into two categories:model-based and deep learning-based methods.Modelbased SA-ISAR methods comprise physical ISAR imaging models based on explicit mathematical formulations.However,due to the high nonconvexity and ill-posedness of the SA-ISAR problem,model-based methods are often ineffective compared with deep learning-based methods.Meanwhile,the performance of the existing deep learning-based methods depends on the quality and quantity of the training data,which are neither sufficient nor precisely labeled in space target SA-ISAR imaging tasks.To address these issues,we propose a metalearning-based SA-ISAR imaging method for space target ISAR imaging tasks.The proposed method comprises two primary modules:the learning-aided alternating minimization module and the metalearningbased optimization module.The learning-aided alternating minimization module retains the explicit ISAR imaging formulations,guaranteeing physical interpretability without data dependency.The metalearning-based optimization module incorporates a non-greedy strategy to enhance convergence performance,ensuring the ability to escape from poor local modes during optimization.Extensive experiments validate that the proposed algorithm demonstrates superior performance,excellent generalization capability,and high efficiency,despite the lack of prior training or access to labeled training samples,compared to existing methods.

关 键 词：逆合成孔径雷达 稀疏孔径ISAR成像 学习辅助 非凸优化 元学习 

分 类 号：TN957.51[电子电信—信号与信息处理]