GAN Imaging Scheme Inspired By Inverse Scattering Model

The problem of electromagnetic inverse scattering is concerned with the reconstruction of geometric information and spatial distribution of electrical properties of the medium object from the measurement scattering field in the imaging space.Solving the electromagnetic inverse scattering problem becomes challenging due to its inherent nonlinearity and morbicity,and its significant enhancement with increasing contrast and decreasing signal to noise ratio(SNR).The traditional optimization method can easily incorporate the inverse scattering domain knowledge,and solve the mismatch between calculated data and measured data by minimizing the objective function,which can iteratively obtain good reconstruction results.However,due to the properties of nonlinear iteration,this method has some common problems,such as how to choose appropriate regularization parameters,sensitive to initial values,high computational cost and poor real-time performance.Although the nonlinear inverse scattering problem can be linearized,the premise of linearization is that the object is a weak scatterer,that is,the difference between the incident field and the total field is small.Therefore,this paper takes problem-driven-model,fusion-network design as the general idea,aiming to reduce the computational cost and improve the quality of inverse scattering imaging under high contrast and low SNR.In this way,a method that combines traditional algorithms,domain knowledge and deep learning is proposed.The details are as follows:Firstly,the inverse scattering problem of two-dimensional transverse magnetic wave polarization is considered in this paper.Through the forward model of electromagnetic field and the Lippmann-Schwinger equation,the equation of state and the data equation about the total field and the scattering field are established.Based on these two basic equations,the solving methods of electromagnetic inverse scattering problem are divided into two types: non-iterative linear method and iterative nonlinear method,and the differences in reconstruction performance of the two kinds of methods are analyzed and compared.Next,although deep learning technology can provide solutions for real-time performance and noise robustness,it is not easy to integrate domain knowledge and mathematics into the input or internal architecture of deep networks.In order to narrow the gap between traditional objective function method and deep learning method and take advantage of both methods,this paper presents some ideas of applying linear and nonlinear inversion techniques and domain knowledge to neural network framework.Finally,we inspired by the model-driven deep learning inversion method,the required neural network architecture was designed by combining physical knowledge and optimization algorithm,and then a generative adversarial network imaging method based on inverse scattering physical model was proposed.This method maps the iterative updating formula of the traditional algorithm and the physical model of the electromagnetic inverse scattering to the corresponding learnable sub-modules in the generative network,and each sub-module cascaded to form a deep unrolled iterative network.At the same time,the discriminative network was introduced to ensure the fidelity of the reconstructed results.Two networks are trained alternately by the generative adversarial learning strategy to reconstruct the spatial distribution of the medium target.On this basis,pretraining model and perceptual loss function are introduced to meet the requirements of high quality and high precision imaging in harsh environments.