Research On Deep Learning Based Electromagnetic-scattering Problems And Electromagnetic Inverse-scattering Problems

With the development of science and computer technology,microwave imaging technology is widely used in some areas,such as biomedical imaging,geophysics,remote sensing and so on.In fact,the physical process of electromagnetic inversion imaging includes the electromagnetic scattering problems and the inverse scattering problems,i.e.,the forward problem and the inverse problem of electromagnetic imaging.The large-scale matrix inversion is involved in solving electromagnetic scattering problem by traditional methods,which will cost huge computing resources.And when solving the inverse electromagnetic scattering problem,it will inevitably encounter two major challenges: morbidity and non-linearity,which cost huge computing resources and affect the quality of imaging at the same time.Therefore,how to improve the efficiency of solving methods in calculating the electromagnetic scattering as well as the inverse scattering problems effectively,and improve the quality of image are the main research contents of this dissertation.Firstly,in order to solve the problem that the huge computing resources cost in calculating scattering fields by traditional methods,a fast method based on Generative Adversarial Networks(GAN)is proposed.As the traditional algorithm involves the large-scale matrix inversion and consumes huge computing resources when calculating the forward induced current,a forward induced current learning method(FICLM)is proposed to solve it.The model of the network is Pix2pix-Gan,of which the input information is the known contrast and the incident field.Through the special design of the model,this method can predict the induced current even though the incident antenna is at any angle.The test samples show that the model has a strong generalization ability and can predict the induced current of some complex scatterers.Under the same grid division,the induced current obtained by this method is more accurate than that of the traditional Mo M,and the calculation speed is more than 20 times faster than the method of moments(Mo M).And with the increase of the scatterers' contrasts,the computational speed advantage becomes more obvious.Secondly,a static gesture recognizer based on scattered field data and complex CNN network is proposed.This proposed network is different from the network based on images,and it is a complex CNN network added the attention module.The results show that both of them can effectively improve the recognition accuracy.At the same time,the scattered field data is smaller than the image data and has more advantages in computer memory.Finally,under the background of new integral equation(NIE)and the research algorithm of Fourier bases-expansion(FBE)based on the contrast source,this dissertation proposes a physics-based and learning-based method to solve the highly nonlinear inverse scattering problems.The network is Pix2pix-Gan added the selfattention module.The test results show that this method can effectively solve the highly nonlinear inverse scattering problems,and the whole calculation process only costs about 1 second.