Research On Real-Time Microwave Imaging Method Based On Deep Learning

The electromagnetic inverse scattering problems(ISPs)contain two parts:electromagnetic forward calculation and electromagnetic inverse calculation.Microwave imaging is one of the electromagnetic inverse calculation problems.With the development of science and technology,microwave imaging technology has been widely used in many fields,such as oil exploration,satellite remote sensing,biomedical detection and so on.Whereas,there are still two major challenges in ISPs:nonlinearity and ill-posedness,which will not only lead to high computational cost,but also affect the final image quality.The main topic of this paper is how to use neural network to improve the speed and accuracy of electromagnetic inversion.Firstly,due to the problem of unsatisfactory end-to-end inversion results,this paper presents an adjustable scalable cascaded convolution neural network(SC-CNN),which can achieve fast,high quality end-to-end inversion,in which the neural network is taken to directly build a mapping between the scattering field and the high-resolution images,the design of multi-resolution labels makes more physical characteristics to be preserved in the imaging process.It alleviates the black-box effect of the network to some extent.By comparing the test results of multiple data sets,it is concluded that the SC-CNNs have better generalization ability.And compared with some existing schemes,both computational efficiency and inversion accuracy are improved.The number of cascade modules can be scalable according to the difficulty of inverse problems,it can be added or reduced the corresponding blocks to obtain better results.Secondly,in view of some cases where data labels are damaged or difficult to get,this paper proposes a microwave imaging method based on Semi-Supervised Learning(SSL).In SSL,there are both labeled data and unlabeled data.And when the amount of labeled data is small,unlabeled data can be used to expand the number of samples,to obtain better inversion results.The test results show that SSL can improve the training results,whether in different data sets or in the case of different proportions of two kinds of data.Finally,this paper presents a limited-aperture data imaging method based on Generative Adversarial Network(GAN).The GAN is used to retrieval the limitedaperture data to the full-aperture data,and then perform simple processing on the recovered data.At the end,the low-resolution images obtained are brought in the trained Unet network to get high-resolution images.By testing multiple data sets and two different antennas angles,it is proved that this method can solve the limited-aperture imaging problem to some extent.