Research On The Key Technology Of Electromagnetic Inverse Scattering Based On Deep Learning

The inverse scattering imaging method based on deep learning is researched from two aspects in this thesis.On the one hand,an imaging algorithm based on the complex convolution CUnet model and an imaging algorithm based on the UnetPlus model are proposed from the perspective of improving the network structure.On the other hand,the scattered electric field data set and the pseudo-spectral data set are proposed from the perspective of improving the quality of the training set and combining the traditional inverse scattering method.The main contents are as follows:First,the advantages and disadvantages of traditional inverse scattering imaging technology based on iterative and non-iterative method are summarized.Then the latest progress of deep learning in the field of inverse scattering and the research significance of this thesis are introduced.Secondly,five typical inverse scattering imaging methods including born approximation method,back propagation method,born iteration method,contrast source method and subspace optimization method are given from the perspective of the principle of inverse scattering imaging.The calculation principle of convolutional neural network and other convolutional networks such as convolutional neural network and fully convolutional neural network,and three existing inverse scattering imaging methods based on Unet model are introduced from the perspective of deep learning network structure model.In order to solve the problem that the real or imaginary part must be discarded in the real Unet model,an inverse scattering imaging method based on the complex Unet model is proposed.The core of complex Unet is to use real and real parts,real and imaginary parts,imaginary and real parts,imaginary and imaginary parts calculation instead of real convolution calculation.In addition,function layers such as normalization of complex values,initialization of complex values and node deactivation can be added to improve imaging quality.To further improve network performance and reduce noise causes the problem of unsatisfactory imaging results,an inverse scattering imaging method based on UnetPlus model is proposed.The UnetPlus model uses a nested convolution structure instead of a connection layer structure.The data features of Unet's encoding and decoding have large deviations.The nested convolution structure is modified by multiple convolutions and activation functions tocontinuously fit the differences between the data and reduce the semantic gap between the encoder and decoder feature maps.Then,from the perspective of the combination of electromagnetic algorithms and deep learning,high-quality scattered electric field data sets and pseudo-spectral data sets are proposed.The complex Unet model based on the scattered electric field dataset solves the problem of imaging artifacts and improves the imaging quality.The UnetPlus model based on the pseudo-spectral data set reduces the relative error from 14.234 to5.021,which can restore the outline of the scattering medium,refine the difference between different scattering media and have strong noise robustness.Finally,the content of the full text is summarized,and the future work for the next step is given from five perspective.