SAR Automatic Target Recognition Based On Image Super-Resolution Technology

Synthetic Aperture Radar(SAR)Automatic Target Recognition(ATR)is the process of extracting target features from SAR images to determine its category attributes.It can be used in many aspects such as battlefield reconnaissance,military strikes,resource surveys and weather prediction,therefore,SAR ATR has important military value and civilian value.In recent years,SAR ATR technology has attracted the attention of many experts and scholars at home and abroad,and is a hot topic in the field of SAR research.In the SAR ATR process,the resolution of the image has a crucial influence on the recognition result.Low-resolution SAR images are not conducive to feature extraction and target classification due to problems such as unclear target texture and blurred edge contours.However,the high-resolution SAR image acquisition cost is high and depends on the technological breakthrough of the radar imaging system.Therefore,how to improve the recognition accuracy of the low-resolution SAR image without relying on the upgrade of the hardware system is a major problem for SAR ATR.Since the introduction of image super-resolution technology,it has played an important role in battlefield reconnaissance,satellite remote sensing imaging,medical imaging and media digital equipment.This thesis combined with the military application needs,with the support of the 13th Five-Year Plan advance research and Youth Program of National Natural Science Foundation of China,a SAR automatic target recognition method based on image super-resolution technology was carried out,focusing on the single SAR image super-resolution algorithm and SAR image classification method based on deep learning.The main research contents include:1.Basic theory of image super-resolution and basic method of SAR ATR.In order to reduce the negative impact of background noise and target shadow on SAR ATR,and improve the recognition accuracy,it is necessary to preprocess the SAR image,and then extract the target area accurately.First,the basic problems in the field of image super-resolution are explained,and the SAR image is reconstructed using traditional interpolation algorithms such as Bicubic and Super-Resolution Convolutional Neural Network(SRCNN).Secondly,the classic SAR ATR method--template matching method and SAR image classification method based on Support Vector Machines(SVM)are introduced,and the basic processing flow of SAR ATR is described in detail.Finally,SAR image preprocessing was performed on the publicly moving and stationary target acquisition and recognition(MSTAR)data,and the common image super-resolution methods and traditional SAR image classification methods were simulated to provide theoretical and technical support for S AR image super-resolution enhancement and SAR target accurate recognition in next work.2.Aiming at the problems of gray discontinuity of reconstructed images obtained by traditional interpolation algorithms and SRCNN and the serious lack of high-frequency information,a SAR image enhancement method based on Enhanced Deep Residual Networks(EDSR)is proposed.First,use the downsampling method to generate "low-resolution-high-resolution"image sample pairs;second,input the image samples into the EDSR network to fully train the mapping relationship between low-resolution images and high-resolution images,and extract the image high-frequency detail features,reconstruct high-resolution SAR images;then,input the reconstructed images into the trained EDSR network to enhance the SAR images.Finally,the MSTAR data was used for simulation verification to achieve high-resolution reconstruction of SAR images.3.Aiming at the challenge that EDSR cannot accurately recover the texture and edge information of the target surface and cause artifacts,a SAR image enhancement method based on Super-Resolution Generative Adversarial Network(SRGAN)is proposed.First,input different types of "low-resolution-high-resolution" image sample pairs into SRGAN separately.Since SRGAN adds an adversarial network structure,the loss function is composed of content loss and adversarial loss,and the high-resolution images are more fully learned the spatial features of the image,thereby achieving high-resolution reconstruction of the texture and edge information of the low-resolution SAR image and obtain a reconstruction model with category attributes;secondly,input the test sample into the reconstruction model to obtain the enhanced SAR image;finally,the simulation using MSTAR data verified that SRGAN accurately enhanced the high-frequency information of SAR images,which laid the foundation for the subsequent accurate identification of SAR targets.4.Aiming at the problems that traditional classification methods rely on prior information and manual extraction of separable features is time-consuming and labor-intensive,this chapter uses Deep Convolutional Neural Network(Deep Convolutional Neural Network,DCNN)to classify the enhanced SAR images.First,build a DCNN structure framework based on Visual Geometry Group(VGG);then,the enhanced SAR image obtained based on SRGAN is divided into a training set and a test set,and send the training set into the DCNN for recognition training,extract the deep identifiable features of the target in the SAR image;finally,input the test set into the trained DCNN model to obtain the final recognition result.At the same time,the effectiveness and generalization ability of the proposed method were verified using the MSTAR dataset under Standard Operating Condition(SOC)and Extended Operating Conditions(EOC).