Research On ISAR Resolution Enhancement Based On Deep Learning

ISAR has become a common means to acquire high-resolution radar images of noncooperative moving targets in recent years.High-resolution radar images can represent more target details.In order to solve the problems of traditional methods in ISAR resolution enhancement,this paper focuses on deep learning.The main research work and results are shown as follow:The first chapter is introduction.It reviews the research status of ISAR resolution enhancement and analyses the shortcomings of traditional methods.Many breakthroughs in image super-resolution task based on deep learning are summarized.The second chapter describes concrete implication of ISAR resolution enhancement based on ISAR imaging theory.And ISAR resolution enhancement index is proposed for evaluating ISAR resolution enhancement performance.The third chapter proves the feasibility of improving ISAR resolution by deep learning,and proposes the framework of ISAR resolution enhancement using deep learning.The mean square error(MSE)between low-resolution ISAR image and its corresponding high-resolution ISAR image is selected as loss function of deep residual network to learn the end-to-end mapping between them.The trained deep residual network is adopted as the reconstruction model of high-resolution ISAR images.For analyzing the performance of this model,we study the effect of echo signal-to-noise ratio and residual network architecture on resolution enhancement.In addition,we analysis the features extracted by this residual network.Compared to traditional methods,this method can simultaneously enhance range and azimuth resolution without model estimation.And this method can effectively avoid false points so that it has higher estimation accuracy.The fourth chapter enhancing ISAR resolution by a generative adversarial network.The weighted sum of absolute loss and adversarial loss is used as loss function of generator to solve the limited enhancing factor and inaccuracy recovery of weak point scatters caused by MSE.Experimental results show this method can achieve higher enhancing factor than other methods based on neural networks,and proves that weak point scatters can still be recovered accurately under low signal-to-noise ratio.The fifth chapter summaries the research contents and innovations of this thesis.And we further discuss the advantages and limitations of these ISAR resolution enhancement methods based on deep learning.