Research On Deep Learning Based ISAR Sparse Imaging Technology

Inverse Synthetic Aperture Radar(ISAR),as a radar imaging technology to obtain shape and structure information of non-cooperative targets,has important research and application value in the fields of strategic early warning and space surveillance.Affected by fierce attack and defense confrontation,electronic interference and other factors,the received echo of ISAR usually has under-sampling and unevenness,resulting in sparse aperture.For sparse aperture ISAR imaging,traditional imaging algorithms will face problems such as high side lobes,strong grating lobes,and reduced resolution.The sparse imaging algorithm based on compressed sensing(CS)can overcome the above shortcomings of the traditional ISAR imaging algorithm and achieve high-precision imaging of sparse aperture ISAR,but it also introduces a series of new problems such as high computational complexity and poor parameter adaptation.In order to realize the high-efficiency and high-precision imaging of sparse aperture ISAR,this thesis combines deep network and CS sparse imaging algorithm to carry out research on ISAR sparse imaging method based on deep learning.The main work and innovations of this thesis are summarized as follows:1.The basic principles of ISAR imaging are studied,the geomety and signal model of ISAR sparse imaging are established,and typical ISAR sparse imaging algorithms are introduced.The solution steps of ISAR sparse autofocusing and imaging methods are analyzed to provide a theoretical basis for the research of high-resolution ISAR sparse imaging technology.2.The deep learning-based ISAR sparse imaging technology is studied,which combines the convex relaxation sparse reconstruction algorithm and the learning ability of deep networks.Firstly,in view of the flexibility of sparse representation,a nonlinear transformation based on convolution and activation function is proposed,and combined with the deep unrolling network based on ISTA,a convolution iterative soft threshold(CIST)network is established,which improves the flexibility and convergence speed of the ISTA.Secondly,for the complex-valued processing problem of deep networks,a Hierarchical Channel-Mixed Framework(HCMF)is proposedto achieve efficient estimation of network parameters by selectively decomposing and combining complex values in the network structure;by deep unrolling the AMP algorithm,HCM-AMPN network is obtained,and the advantages of HCM-AMPN in imaging performance and imaging speed are confirmed through the theoretical derivation and experiments.3.The deep learning-based sparse aperture ISAR autofocusing imaging method is studied.Firstly,combined with the feature learning of convolutional neural network and the defocusing characteristics of ISAR images,the ISAR sparse autofocus method(AFUNet)based on U-Net network deployment is studied;secondly,combined with the minimum mean square error criterion and the AMP sparse imaging algorithm,the phase estimation compensation process is integrated into the CS solution framework,and an efficient and high-precision ISAR sparse autofocusing algorithm(AF-AMP)is studied.The ISAR sparse autofocusing imaging network(AF-AMPNet)further improves the quality and efficiency of ISAR sparse imaging through layer-by-layer adaptive parameter optimization and high parallelization of the deep network.Simulation and measured data verify the effectiveness and superiority of the above autofocusing method.