Research On Self-supervised Learning Based Electromagnetic Scattering And Inverse Scattering Problems

Electromagnetic inversion imaging is a physical process based on electromagnetic wave scattering and inverse scattering,which includes both forward and inverse problems.The forward problem involves calculating the scattered waves on an object given its electromagnetic properties and incident wave,using the Maxwell equations to predict the propagation of electromagnetic waves inside the object.The inverse problem,on the other hand,involves inferring the electromagnetic properties of the object given the scattered waves.Solving the electromagnetic scattering problem typically involves numerical simulation,which requires significant computational costs.The complexity and challenges of the electromagnetic inverse scattering problem arise from its nonlinearity and ill-posedness.Traditional inversion methods are based on mathematical models established on physical principles,where an objective function is optimized to resolve the mismatch between estimated and measured data,iteratively obtaining the characteristics of the scatterer.Although traditional inversion methods allow for the incorporation of knowledge in the inverse scattering field and have good physical interpretability,they also suffer from common problems such as sensitivity to initial values,high computational cost,and poor real-time performance.Therefore,this article focuses on how to use deep learning techniques to overcome the problems in solving electromagnetic scattering and inverse scattering problems.When the size and shape of the scatterer become complex,the number of unknowns to be solved also increases,requiring the inversion of a dense matrix.This leads to a significant increase in the required computational resources and time cost in traditional methods for computing the scattering field.To address this,a fast electromagnetic scattering problem-solving method is proposed.The idea is to divide the total incident field into two non-overlapping incident field components and use a self-supervised generative adversarial network to construct a mapping relationship between the known incident field components and the induced currents with contrast.In addition,an auxiliary self-supervised mechanism based on the rotation loss of the induced current corresponding image domain is added to stabilize the mapping process in the network.The results show that,under the same grid partition,this method has better accuracy and speed in computing scattering field data than traditional methods.Although integrating the field knowledge and its updating formula into the internal architecture of the neural network can achieve good reconstruction performance,only a few physical variables are explicitly constrained to meet data consistency.To address this,an inversion method based on an inverse scattering model is proposed.This method draws on the idea of self-supervised learning strategy,incorporating the distribution information of the scattering field into the network,and combining contrast to jointly constrain the loss function in the training stage to refine the solution space.At the same time,an optimization network with self-attention mechanism is introduced to suppress poor feature elements and improve reconstruction performance.The test results show that the proposed method outperforms traditional and some learning-based methods in terms of reconstruction accuracy,reconstruction time,and computational complexity.