High-Resolution Synthetic Aperture Radar Imaging Based On Deep Learning

Synthetic Aperture Radar(SAR)is an active microwave remote sensing system with all-time and all-weather observation capabilities.With the increase of military and civilian demands,higher requirements are put forward for the resolution of SAR images.Although increasing the bandwidth of the transmitted signal and the length of the synthetic aperture can improve the theoretical resolution of the SAR system,the hardware development cycle is long,the cost is high,and the development of related technologies is limited.Utilizing digital signal processing technology to improve the resolution of SAR images is a low-cost and feasible way,and it's also a hot topic and difficult problem of current research.SAR autofocus and sparse imaging are two key technologies in high-resolution SAR imaging processing.The existing SAR autofocus methods cannot be well balanced in terms of quality and speed,and some methods are highly dependent on strong scattering points and are not universally applicable.The iterative reconstruction process of traditional SAR imaging methods based on compressed sensing is time-consuming and poor in robustness.In addition,the existing SAR compression imaging methods based on deep learning cannot quickly adapt to new observation data.In order to overcome the above problems,several novel technologies and methods for high-resolution SAR imaging based on deep learning are proposed in this thesis.The main research contents and results are as follows:1)A deep learning-based framework and two algorithms for SAR autofocus are designed.The traditional minimum entropy-based algorithms(MEA)need hundreds of iterations and are computationally expensive.To overcome this,a deep learning-based SAR autofocus framework is proposed.In this framework,deep neural networks are applied for feature extraction and phase error estimation from the defocused SAR images.The entropy metric is used as the loss.After training,the network learned the rules of autofocus from a large number of examples and can be utilized to focus new SAR images,which greatly reduces the focusing processing time.Firstly,based on the framework,an AutoFocus network(AFnet)based on a deep convolutional neural network is proposed.Secondly,in order to further improve the performance,a Progressive AutoFocus network(PAFnet)model is proposed to predict the phase error coefficients progressively.The complexity of the problem and the learning difficulty of the network is reduced.Thus,the estimation accuracy and image focusing quality are improved.Experimental results show that,compared with the traditional traditional minimum entropy-based methods,the focusing speed of the proposed AFnet method on E5-2696v3 CPU and 1080ti GPU are increased by 6?35 and 11?232 respectively.Furthermore,compared with AFnet,the speed of the proposed PAFnet method on E5-2696v3 CPU and 1080ti GPU are increased by 35%?40%and 23%?30%respectively.2)A SAR autofocus algorithm based on ensemble Convolutional Extreme Learning Machines(CELMs)is designed.Since the training progress of the deep learning-based autofocus methods is very time-consuming,two SAR autofocus methods with fast training speed are proposed based on CELMs and ensemble learning.Multiple weak SAR auto-focusers based on CELMs are integrated to improve the overall focusing performance.Firstly,the traditional extreme learning machine is extended to a convolution extreme learning machine to adapt to the image task.Secondly,a new metric-based combination strategy is proposed.Finally,based on bagging and divide-conquer ensemble learning methods,two autofocus algorithm models are proposed.Experimental results on real data show that the training time of the proposed method is less than 2%of that of the deep learning method.3)A two-dimensional finely SAR autofocus algorithm based on complex-valued convolutional neural networks is designed.To deal with the two-dimensional autofocus problem of complex SAR images,complex-valued convolutional blocks are designed.The coarse focused SAR image is utilized as the training sample to learn the complex-valued filter.The amplitude loss and Fourier loss in the image domain are introduced to ensure the consistency between the output data and the real data.Entropy loss is applied to further improve the image autofocus quality.Experimental results show that this method is superior to the traditional methods in autofocus quality(entropy,contrast)and speed.4)A compressed sensing SAR imaging method based on deep learning and meta-learning is designed.Traditional compressed sensing-based SAR imaging methods have the following disadvantages:very time-consuming,sensitive to imaging parameters,and poor robustness.To overcome this,we propose a robust SAR compressive imaging method based on deep networks,which utilized a network to learn the inverse mapping from echo to scene.This method does not require precise parameter information and is non-iterative.Only a single network forward propagation operation is required to reconstruct the scene.To further improve the performance of the designed network model.A compressed sensing SAR imaging method based on meta-learning is proposed.The Model-Agnostic Meta-Learning(MAML)method is applied to learn the solution of the model from a large amount of observation task data,which can quickly adapt to the new task.Experimental results show that under the same sampling rate,compared with the traditional compressed sensing-based methods,this method has a faster reconstruction speed and a lower noise level.When faced with new task data,it only takes a few iterations for the network to converge.5)A sparse SAR imaging algorithm based on a complex-valued iterative shrinking threshold algorithm network is designed.Aiming at the problem of high side-lobe and noise-level of SAR image reconstructed by traditional matched filtering algorithms,a sparse SAR imaging method based on a deep unrolled network is proposed.Firstly,the original real-valued Iterative Shrinkage Threshold Algorithm(ISTA)algorithm is extended to the complex-valued version(CISTA).Secondly,let the image degradation model and the parameters of CISTA be learnable.Finally,the CISTA optimization iterative process is deeply unrolled to form a deep sparse SAR imaging network(CISTAnet).Experimental results show that,compared with the sparse SAR imaging method based on ISTA and the matched filter method,the reconstructed image by CISTAnet has a higher target background ratio,and can finish autofocus while sparse imaging.