Research On SAR Autofocus Deep Convolutional Neural Network

Synthetic Aperture Radar(SAR)has an irreplaceable position in the military and civilian fields due to its unique advantages such as all-day,all-weather and twodimensional high resolution.With the wide application of SAR in various fields,higher requirements are also placed on the quality of SAR images.In this paper,a method based on deep convolutional neural network is proposed for the two problems of SAR image autofocus and SAR moving target autofocus to improve the quality of SAR images.advantages of the network.The main content of this article is divided into three parts:Firstly,a stationary target imaging model is established,and the Range Doppler(RD)imaging algorithm for stationary targets is deduced.The paper gives the RD simulation results of correct imaging,imaging with phase error and imaging with velocity error of SAR platform.Then the geometric model of moving target imaging is established,the results of RD imaging of moving targets with clutter parameters are deduced,the influence of range and azimuth velocity on the imaging results of moving targets is analyzed,and 1D defocusing is given.and 2D defocused moving point target simulation results.Finally,the basic concepts of the convolutional neural network module are introduced.Secondly,to solve the problem of blurring of SAR images due to phase error or velocity error,a autofocusing method for SAR images based on residual block Unet network is presented.By generating SAR images with different degrees of phase error and velocity error,a part of the scene image is used for the training of the network,and the remaining part of the scene is used for the test of the network.The final trained network defocuses SAR images on scenes and defocus levels that the network has never seen before during training.The experimental results of the final SAR image autofocusing show that the residual-based Unet network proposed in this paper is better than the traditional minimum entropy algorithm and the general Unet network for focusing on SAR images.Finally,in order to solve the problem of one-dimensional or 2D defocusing of SAR moving targets due to the existence of range and azimuth velocity in SAR images,a residual block-based Unet network autofocus method for SAR moving targets is proposed.The moving target images with different scattering intensities,different defocusing directions and degrees of defocusing are simulated by RD imaging algorithm,and Gaussian white noise of different degrees is added to obtain simulated defocused moving targets with different signal-to-noise ratios.These simulated defocused samples are used for training of five deep convolutional neural networks of Unet,improved Unet,residual convolutional neural network,generative adversarial network,and Res-Unet proposed in this paper.Finally,the trained network is tested for focusing on the simulated and measured defocused SAR moving target images,and the focusing results are evaluated through three dimensions.The experimental results of the final SAR moving target autofocusing show that the Res-Unet proposed in this paper is better than the other four deep convolutional neural networks in focusing on SAR moving targets.