| By expanding the synthetic aperture dimension,three-dimensional synthetic aperture radar(3D SAR)is capable of revealing the 3D scattering distribution of targets,showing great potential in many civil and military fields,such as topographic mapping,resource investigation,reconnaissance and early warning.Compressed sensing(CS)breaks the limitation of the Nyquist sampling theory and is capable of accurately reconstructing sparse signals from incomplete measurements,showing high potential for simplifying 3D-SAR system and improving imaging accuracy.However,the current CS-driven imaging algorithms are limited by time-consuming computations,cumbersome parameters’ debugging,and poor scenario adaptability.The self-learning signal reconstruction framework maps iterative reconstruction algorithms to deep networks,which can significantly improve the convergence speed,computational efficiency,and reconstruction accuracy of the corresponding iterative algorithms.Based on the self-learning signal reconstruction framework,this dissertation focuses on key issues such as 3D-SAR inverse problem modeling,network mapping,and motion-error compensation.The main contents are as follows.1.A vectorization-based 3D-SAR sparse learning imaging framework is proposed.By analyzing the spatial-temporal scattering characteristics of 3D scenes and constructing the self-learning mechanism of sparse representation of scenes,the 3D-SAR sparse selflearning imaging networks based on matrix-vector model are designed which improve the imaging accuracy,convergence speed,and computational efficiency.2.A kernel-based 3D-SAR sparse learning imaging framework is proposed.By investigating the linear approximation mechanism of 3D-SAR signal model,the kernelbased sparse imaging model is established.Then,the range migration sparse reconstruction network,lightweight FISTA-inspired sparse reconstruction network,and the efficient ADMM framework with single-frequency holographic operator are designed,which significantly reduce the computational complexity of near-field 3D-SAR sparse imaging and break through the reconfigurable scale limitation of large-scale imaging problems.3.3D-SAR learning imaging methods for weakly/non sparse scenes are proposed.By analyzing the priori information characteristics of weakly sparse scenes,the enhanced imaging models based on priors including joint sparse and low-rank,complex-valued Total Variation(TV),and learned self-adaptive sparse are established.The data-driven Imaging network with learned low-rankness and sparsity,complex-valued TV-driven network with nested topology,and perceptual learning imaging framework are designed,which improve both accuracy and efficiency in reconstructing 3D images form incomplete echoes in weakly/non sparse scenes.4.The learned 3D-SAR sparse autofocusing method is proposed.By investigating the coupling characteristics of the slant-range error components,the decoupling approximations of position error along different axis are derived.Then,a batch-wise autofocusing mechanism is established,based on which the learned autofocusing network is designed to reconstruct well-focused 3-D SAR images from incomplete and error-polluted echoes.In conclusion,this dissertation systematically establishes the theoretical framework of 3D-SAR sparse learning imaging.Besides,a variety of novel and effective learning imaging algorithms are proposed,whose effectiveness is validated by 3D-SAR simulations and measured data tests,providing an important theoretical guidance and technical foundation for the research and application of 3D-SAR learning imaging techniques. |
