As a high-resolution imaging radar technology that directly acquires the rich information of target 3D spatial distribution,scattering characteristics and texture structure,millimeter-wave three-dimensional synthetic aperture radar(3D-SAR)has important research significance and application value in military and civil fields such as terrain mapping,autonomous driving and battlefield sensing.Due to the influence of system hardware,data processing,error interference and other factors,the traditional SAR imaging theory and method are limited by the resolution Rayleigh criterion and face the problems of low resolution,high side flap and gate flap,and it is difficult to obtain high quality millimeter wave 3D-SAR images under sparse sampling.The sparse imaging algorithm based on compressed sensing(CS)uses iterative optimization solution to achieve accurate reconstruction of the observed target under sparse undersampling,which can effectively overcome the problems of traditional imaging methods.However,the CS method has new problems such as high computational complexity and difficult parameter tuning,which limit its application to millimeter-wave 3D-SAR fast imaging.In order to realize millimeter-wave 3D-SAR imaging with high efficiency and high accuracy,fusing CS reconstruction theory and deep learning methods,this thesis carries out research on learning-driven millimeter-wave 3D-SAR 3D imaging methods based on the following main works:1.The mechanism of millimeter-wave 3D-SAR imaging is studied.The theoretical basis and process of the classical frequency domain fast imaging algorithm and CS sparse imaging algorithm are introduced,the millimeter-wave 3D-SAR ground imaging experimental system is designed,the high-resolution 3D millimeter-wave radar dataset(3DRIED)is constructed,and three classical algorithms for imaging processing were completed,filling the gap of the dataset in the field of 3D-SAR and providing data basis for the research of its efficient and high-precision imaging methods.2.A learning-driven millimeter-wave 3D-SAR imaging method is investigated.Firstly,a joint imaging framework of frequency domain fast imaging operator and CS sparse reconstruction is designed to avoid large-scale matrix operations and improve imaging efficiency.Secondly,the low-rank characteristics of millimeter-wave 3D-SAR echo data are studied and a sparse 3D imaging method based on low-rank reconstruction(LSIIA)is designed.Finally,for the network characterization processing of complex-valued echo signals,a sparse 3D imaging method based on split iterative network expansion(SISR-Net)is proposed for the network characterization processing of complex-valued echo signals,which solves the problems of difficult complex-valued signal network characterization and poor adaptive iterative parameters,and improves the efficiency and robustness of millimeter-wave 3D-SAR imaging.The imaging advantages of SISR-Net,such as high efficiency,high accuracy and strong generalization,are verified by multiple sets of experiments.3.A learning-driven millimeter-wave 3D-SAR autofocusing imaging method is investigated.Firstly,an iteratively optimized sparse autofocusing imaging method(AFKAMP)is designed jointly with an approximate message passing(AMP)imaging algorithm and an image quality optimality criterion.Secondly,based on the AF-KAMP iterative estimation framework and deep unfolding network,a autofocusing model-driven sparse imaging network(SAF-3DNet)is proposed to improve the accuracy and efficiency of millimeter-wave 3D-SAR autofocusing imaging through hierarchical optimized parameter learning and efficient parallel processing of deep networks.The effectiveness of SAF-3DNet is verified through simulation and real data imaging experiments. |