Sparse Aperture ISAR Imaging Based On Deep Learning

Inverse Synthetic Aperture Radar（ISAR）has become a common method to obtain high resolution radar images of non-cooperative moving targets in recent years,and has been put into various civil and military fields.However,some radar echos will miss due to target maneuvering and strong noise interference in practical application,which makes the traditional RD imaging algorithm invalid.For incomplete radar echo,the traditional RD algorithm fails to produce satisfactory ISAR images,in which case,the ISAR image reconstruction is generally completed by compressed sensing algorithm.This method usually adds sparse priors to the model through ₁l regular term constraints,but the prior model based on regular term constraints is not the ideal prior model.To solve this problem in sparse aperture ISAR imaging,this paper proposes a sparse aperture ISAR imaging algorithm based on Pn P-ADMM by combining traditional compressed sensing algorithm with deep learning and learning the prior structure information of the target from the data through neural network.However,due to disadvantage of iteratively calling the same denoising network in Pn P framework.We further improve the limitations of sparse aperture ISAR imaging using deep learning in the Pn P framework and propose the U-ADMMNet imaging algorithm.Finally,an ISAR image enhancement algorithm based on CBAM-Unet+algorithm is proposed for ISAR images after imaging.The research work and content arrangement of this paper are as follows:The first chapter is the introduction,which introduces the research background and significance of this topic.The development of sparse aperture ISAR imaging technology based on compressed sensing,deep learning,deep unrolling and plug-and-play are analyzed and the experimental data used in this paper are introduced.In the second chapter,the radar target scattering model,the basic principle of ISAR imaging and the ISAR imaging model under sparse aperture are firstly introduced.Then,the effects of different sparse modes on imaging results are analyzed through simulation data and measured data.At the same time,the evaluation indicators used in the evaluation of imaging results are introduced.In the third chapter,aiming at the limitations of the prior model of traditional ADMM algorithm,Pn P framework and ADMM algorithm are combined to conduct sparse aperture ISAR imaging,and an imaging method of Pn P-ADMM is proposed.The structural prior information learned by deep network from training data is used to replace the ₁l norm constrained priors of traditional ADMM algorithm in this method,and a more accurate prior model will be obtained.This chapter first introduces the derivation process of sparse aperture ISAR imaging algorithm based on ADMM,and deduces the process of Pn P-ADMM imaging algorithm.The experimental results of electromagnetic simulation data and measured data verify that the proposed method is superior to the traditional ADMM algorithm in imaging quality.In the chapter 4,based on chapter 3,aiming at the limitations of the current Pn P framework based on deep learning to call the same denoising prior network in algorithm iteration,a new U-ADMMNet imaging algorithm which combines the advantages of Pn P framework and deep unfolding method is proposed.Firstly,the ADMM algorithm is unrolled into a network by deep unfolding method,and then the complex Unet network is replaced by the denoising prior module in the ADMM network to form the proposed U-ADMMNet network,which can call the denoising prior network with different parameters in different iterations of the algorithm.Finally,the simulation and measured experimental data verify that the proposed method is superior to ADMM and CRL1-ADMM in image evaluation indexes such as correlation coefficient,root mean square error and peak signal-to-noise ratio.Chapter 5 proposes to introduce full scale connection and CBAM attention module in UNet to solve the problem that Unet does not make full use of radar image semantic information in ISAR image enhancement task.The detail and structure information of the image are captured from the whole scale,and the channel and spatial attention module are introduced into the multi-scale feature fusion layer to enhance the feature expression ability of the network.Simulation and experimental results show that the image enhancement effect of the proposed network is better than that of the current deep neural network.The last chapter summarizes the work and innovation of the whole paper,and analyzes the limitation of the current work and possible solutions in the future.