Deep Iterative Network Scheme For Born-type Nonlinear Inverse Scattering Imaging

The problem of electromagnetic inverse scattering is the process of reconstructing the medium target through the numerical calculation of the forward model from the measurement data obtained from the receiving antenna.Because of its non-touch and non-destructive characteristics,it is widely used in ground penetrating radar,biomedical imaging,geophysical imaging and other fields.In view of the inherent nonlinearity and morbidity of electromagnetic inversion problems,nonlinear iterative algorithms are usually used for reconstruction,but there are problems such as the need to artificially set the regularization parameters and functions,high computational cost,and easy to fall into local minimums..In this paper,the general framework of Born iteration based on the Born iteration method and the deep iterative network are combined to design a set of regularization functions and parameter-learnable inversion network models,so as to achieve the noise robustness and robustness of the electromagnetic inverse scattering algorithm.The compromise between imaging effect and convergence speed enables real-time reconstruction and high-precision imaging.The specific research content is launched from the following three aspects:First,according to the forward propagation model of electromagnetic waves and the Lippmann-Schwinger equation/integral equation,two basic integral equations of the total field and the scattered field are constructed.Starting with two integral equations,the advantages and disadvantages and differences between the Born iteration method and its related extension methods(such as the variational Born iteration method,etc.)are analyzed.On this basis,a new total field update method based on the traditional Born method based on the principle of subspace is introduced,so that the new Born method has a wider scope of application and a better reconstruction effect.Next,in view of the contradiction between the noise robustness and the imaging effect and the algorithm convergence speed in the Born-type traditional general framework,although the balance factor can be used to solve the problem,there is a certain amount of man-made setting for the value of the balance factor.factor.At the same time,most of the traditional methods use conjugate gradient descent/least squares methods to solve the electromagnetic inverse scattering imaging problem,which to a certain extent exacerbates the reduction in the scope of the algorithm and the increase in the calculation cost of the algorithm.From the extensive application of deep iterative networks in other imaging fields,it can be seen that this method can learn the optimal values of some artificially set functions and parameters in the physical model framework through the training of certain samples,and finally achieve the height of the unknown target scatterer.Precision imaging effect.Finally,based on the above problems,this paper innovatively integrates the deep iterative network with the Born-type general framework,expands the iterative optimization process into a deep iterative network form,and maps the variables in the iterative algorithm to the learnable parameter variables in the network In order to reduce the computational cost and avoid the limitations of algorithm application caused by human factors.By comparing the traditional method with the network method in the simulation data comparison experiment,it can be known that the latter has a faster and more accurate inversion of the target medium than the former.