| The inverse scattering problem is concerned with retrieving the spatial distribution of electrical properties of dielectric targets from the scattering field data measured in the detection region.Due to the inherent nonlinear and ill-conditioned nature of inverse scattering problem,nonlinear iterative optimization with regularization is usually used to solve the problem.However,these methods have high computational costs and cannot be reconstructed in real time in the imaging process,and they need to manually design regularization functions and select regularization parameters according to experience,so the applicability of the algorithm is affected by human factors.In view of this,this paper introduces deep learning technology to solve the above problems from the inversion scheme based on deep learning assistance,the inversion scheme based on physical assistance and the inversion scheme based on model-driven deep learning.First of all,explores the two kinds of inversion scheme based on depth of learning,the general ideas are the traditional iterative or non iterative method,such as back propagation method and VBIM,reconstruction from scattered field data contrast the spatial distribution of the initial value,implement the data domain to the image domain,and then use the end-to-end deep learning network learning image to the image map,Finally,a more accurate contrast spatial distribution image is reconstructed.Secondly,the inversion scheme of deep learning based on physical assistance is explored.Based on the minimum mean square error(MSE)loss function of BIM deep learning-assisted imaging scheme,a more physical loss function under field joint constraint is introduced in this scheme,and a new loss function is defined to obtain the descending direction of the loss function,so as to update the training model.Finally,two deep learning inversion schemes based on model driven are explored.Inspired by the idea of Neumann network,a BIM based Neumann expansion network(BN-Net)is proposed to reverse the solution matrix into Neumann series expansion.The advantages of the algorithm are that the neural network is used to learn regularization parameters,so as to avoid the dilemma of choosing parameters manually and improve the quality of inversion imaging.In addition,a deep iterative unfolding network(IV-Net)based on VBIM is proposed to solve nonlinear inverse scattering problems to improve the quality of reconstruction.Each iteration algorithm of VBIM is expanded into a layered deep neural network.Each iteration maps to each layer of the neural network,and each subproblem is treated as a module.Therefore,the layers are connected to form a deep network structure.In order to evaluate the performance of the two inversion schemes,several representative tests are carried out.The numerical results show that the two inversion schemes are superior to the traditional ones in terms of inversion speed,accuracy and noise robustness. |
