Electromagnetic Inverse Scattering Imaging Combining Physical Model And Multi-scale Iterative Unfolding Network

For the nonlinear and ill-posedness electromagnetic inverse scattering problem,the nonlinear iterative optimization algorithm with regularization is usually used to solve it.Traditional nonlinear iterative methods not only have high computational complexity,but also require manual selection of regularization schemes,which cannot make good use of limited prior information.The multi-scale iterative method can make up for this shortcoming through an adaptive multi-resolution strategy.Considering that its optimization objective function still relies on the traditional iterative algorithm,this paper organically integrates the traditional multi-scale iterative method with the deep network emerging in recent years.A model-driven network with a multi-resolution mechanism is designed,which balances imaging quality and computational complexity to a certain extent.The research content of this thesis is carried out in the following three aspects:First of all,starting from the electromagnetic field forward scattering model and integral equation,the basic idea of solving the inverse scattering problem is deduced,and three solutions are introduced according to the strength of the multiple scattering effect,and then the universal nonlinear iterative optimization algorithm is introduced.Since the subspace class method can provide better imaging results than the contrast source class method,and can achieve faster convergence,in order to echo the following,this paper focuses on three subspace class optimization methods.In addition,in view of the shortcomings of the traditional multi-scale iterative method,such as poor imaging results and easy to fall into local minima,this paper draws on the related algorithms introduced above,uses a new framework for multi-scale iteration,and proposes two subspace-based multi-scale iterative methods.In addition,this paper also proposes a strategy for inversion using multi-frequency data,which can further alleviate the nonlinearity and ill-posedness of the inverse problem.At last,by combining the advantages of traditional multi-scale iterative methods and deep networks,this paper proposes a multi-scale iterative unfolding network imaging scheme that integrates physical models.Two different multi-scale iterative unfolding networks are designed according to the frequency of the data used.Compared with the single-frequency method,the proposed multi-frequency and multi-scale iterative unfolding network not only considers the multi-frequency data,but also innovatively uses the progressive learning strategy,which further enhances the imaging effect on the original basis.