摘要
针对传统逆合成孔径雷达稀疏成像算法存在参数敏感、收敛速度慢等问题,文章以卷积神经网络的自适应参数学习机制为基础,结合模型驱动网络的物理可解释性,提出了一种ISAR稀疏成像架构——深度增强迭代收缩阈值(Deep Augmented-Iterative Shrinkage Thresholding,DA-IST)网络。首先,DA-IST网络将迭代收缩阈值算法的迭代步骤映射至隐藏层中,不仅能够提高可解释性,而且能够在训练过程中学习最优参数;其次,网络在建模过程中考虑了被忽略的高频分量,提高了重构性能;同时,为了提高网络的鲁棒性,用非线性卷积层替代了线性稀疏变换。实验表明,与传统的模型驱动算法相比,DA-IST网络避免了人工调整参数过程,收敛速度更快,成像质量更高,对特征差异较大的数据具有更好的泛化能力。
Addressing the issues of parameter sensitivity and slow convergence in traditional Inverse Synthetic Aperture Radar(ISAR)sparse imaging algorithms,inspired by the adaptive parameter learning mechanism of convolutional neural networks and combining the physical interpretability of model-driven networks,a new ISAR sparse imaging framework known as the Deep Augmented-Iterative Shrinkage Thresholding(DA-IST)network is proposed.Firstly,the DA-IST network maps the iterative steps of the Iterative Shrinkage Thresholding Algorithm(ISTA)into the hidden layers,which not only improves interpretability but also enables learning optimal parameters during training.Secondly,the network takes into account neglected high-frequency components during modeling,enhancing reconstruction performance.Additionally,to improve the network's robustness,nonlinear convolutional layers are employed to replace linear sparse transformations.Experimental results demonstrate that,compared to traditional model-driven algorithms,the DA-IST network eliminates the need for manual parameter tuning,exhibits faster convergence,produces higher-quality imaging,and possesses better generalization capabilities for data with significant feature differences.
作者
潘之梁
户盼鹤
陈凌峰
苏晓龙
刘振
PAN Zhiliang;HU Panhe;CHEN Lingfeng;SU Xiaolong;LIU Zhen(College of Electronic Science and Technology,National University of Defense Technology,Changsha Hunan 410073,China)
出处
《海军航空大学学报》
2024年第5期603-614,共12页
Journal of Naval Aviation University
基金
国家重点研发计划(2021YFB3100800)
国家自然科学基金(62022091、61921001)
国防科技大学青年自主创新科学基金(ZK21-14、ZK23-18)。
关键词
逆合成孔径雷达
稀疏成像
模型驱动网络
深度学习
inverse synthetic aperture radar
sparse imaging
model driven network
deep learning
