Physical-Inspired Deep Network For Electromagnetic Inverse Scattering Imaging Problem

The inverse electromagnetic scattering problem is to reconstruct the physical properties of the detection area according to the measured scattering field data and some prior information.Because of its non-destructive characteristics,electromagnetic inverse scattering imaging is widely used in military,remote sensing,non-destructive detection and other fields,but the problem is essentially nonlinear and ill conditioned.The traditional iterative optimization method for solving the electromagnetic inverse scattering problem has the advantages of high reconstruction accuracy and ensuring the convergence of the objective function,but there are also problems such as the possibility of falling into the local minimum value,too much manual intervention in regularization functions and parameters,long inversion time and high computational cost.In order to solve these problems,this paper proposes an imaging method which combines electromagnetic inverse scattering with depth learning.It is expected that this method has high anti noise and robustness,relevant parameters can be learned,real-time reconstruction can be realized and computational cost can be reduced.Firstly,the research scene of this paper is two-dimensional transverse magnetic wave problem.In the forward model,according to the propagation process of electromagnetic wave and the physical model of measurement,the nonlinear integral equation is discretized,transformed into matrix formula,and the data of scattering field is obtained by finite difference method/moment method.Secondly,in order to solve the nonlinear problem between the measured scattering field and the electrical performance parameters of the target scatterer due to multiple scattering efects,a linear electromagnetic inverse scattering method is proposed,which combines the physical model driving and the depth network structure.The algorithm first ignores the coupling effect between electromagnets,and then obtains the initial value of dielectric constant.Although the initial value is quite different from the actual value,because the initial value already contains the characteristic information of the target,then using this initial value for network optimization can obtain better reconstruction.But the algorithm is only suitable for weak scatterers.When the target is a strong scatterer,the interaction between the targets can not be ignored,which makes the linear ill conditioned problem turn into a nonlinear ill conditioned problem.For the nonlinear ill posed problem,a nonlinear electromagnetic inverse scattering method is proposed based on the linear method,which combines the physical model driving and the depth network structure.Firstly,the initial value of dielectric constant is obtained according to the data equation of electromagnetic inverse scattering,and then the reconstructed dielectric constant is more accurate through network optimization and constant updating of the total field.Finally,aiming at the problem of electromagnetic backscatter,this paper innovatively integrates the physical model of electromagnetic backscatter with convolution neural network method,and transforms the parameters of network structure into learnable variables,thus reducing the calculation cost and realizing real-time or quasi real-time electromagnetic backscatter imaging.Experimental results show that the method is not only efficient and accurate in imaging,but also robust to noise.