Research On Sparse Imaging Method Of Through-the-wall Radar Under Diffraction Tomography Model

Through-the-Wall radar has become a research hotspot because of its penetrating ability,which can obtain the information of hidden targets without destroying the scene.Its three-dimensional imaging provides azimuth,distance and height information,which satisfies the needs of human behavior characteristics such as bone contours,action postures,and object distribution information in buildings.Considering the problems of large amount of data,long acquisition time and side lobe in 3D imaging algorithm,compressed sensing theory is introduced to sparse imaging.However,there are some problems,such as too much memory needed to build dictionary matrix,the selection of threshold parameters in sparse reconstruction convex optimization algorithm and the influence of wall on target imaging.Therefore,in this paper,we build a sparse model of diffraction tomography and use the dual drive of model and data for learning imaging.The specific research contents are as follows:1.In response to the problem of the inability to image the target because the wall echo covers the target echo,a learning iterative shrinkage threshold algorithm(LISTA)is proposed to separate the echo signals of the wall and the target.Firstly,the low rank sparse model is established,then the iterative shrinkage threshold algorithm is used to solve the sparse solution,and the iterative process is mapped into a LISTA network with a multi-layer neural network structure.Finally,the threshold parameters are learned by data-driven to realize the automatic selection of threshold parameters.Simulation and experimental results verify the influence of threshold parameters on the separation of wall and target echo,and the improvement of separation effect by trained threshold parameters.2.In view of the problem that the traditional sparse model dictionary matrix needs too much memory and the selection of threshold parameters of convex optimization algorithm affects the image quality,a learning approximation message passing(LAMP)network is proposed for 3D learning imaging.This method builds a diffraction tomography sparse model by constructing a fast Fourier transform operator on the diffraction tomography algorithm,avoiding the construction of a dictionary matrix,and then uses a modified approximate message passing algorithm to solve the sparse solution,and maps its iterative process a LAMP network with a multi-layer neural network structure,and finally through data-driven adaptive learning of the adjustable parameters in the multi-layer neural network,so as to achieve three-dimensional learning and imaging.Simulation and experimental data processing results show that this method reduces the memory required by the system,avoids the impact of manual adjustment of parameters on imaging quality,and uses GPU acceleration to reduce the time required for imaging.3.In order to solve the problem of 3D sparse imaging with unknown wall parameters,an autofocusing imaging method with alternating learning of wall parameters and sparse coefficients is proposed.In this method,we fix the wall parameters and update the sparse coefficient by LAMP network.Secondly,we fix the sparse coefficient and continue to update the wall parameters by back propagation algorithm in neural network.Finally,we select the minimum loss function as the optimal wall parameter estimation and autofocus imaging result after several alternate iterations.The simulation results show that the method avoids the Taylor series expansion method to solve the change of wall parameters,and reduces the imaging times and optimization time compared with the global search of wall parameters.