Research On SAR Image Processing Based On Deep Learning

Synthetic Aperture Radar(SAR)has the advantages of long-term stable operation,strong adaptability to different climates and good penetration.After decades of development,due to its unique advantages,SAR imaging technology has been widely used in such fields as terrain remote sensing detection,military investigation and identification,real-time monitoring and prevention of natural disasters,and its emergence has far-reaching significance for the reform and improvement of national defense technology and modern civil range.With the rapid development of deep learning in recent years,the application of deep learning in SAR imaging is more and more frequent and there are many researches on target recognition,denoising and target segmentation in SAR image based on deep learning.This thesis mainly completes the SAR image target detection and segmentation experiment based on deep learning technology.The near-field 3D BP algorithm based on GPU parallel acceleration,imaging performance acceleration and quality optimization,imaging fixed channel error analysis modeling and Various network models are mainly discussed in this thesis.Finally,target detection of SAR imaging is accomplished by means of YOLO and heat map,and target segmentation of SAR image is realized by full convolution network,the main research contents and work are as follows:1.The theoretical basis of Linear Array 3D SAR imaging is described.Firstly,the commonly used linear frequency modulation signal and pulse compression technology in SAR imaging are analyzed.secondly,the target echo signal model is established based on the vertical scanning mode of near-field security inspection imaging.then,the ambiguity function and 3D theoretical resolution of Linear Array 3D SAR imaging are given.finally,the commonly used BP algorithm in SAR imaging are analyzed.2.Near-field three dimensional security inspection imaging based on measured data was realized and relevant theoretical analysis was carried out in this thesis.Firstly,the security inspection imaging theory is briefly introduced,mainly including the system of imaging model,maximum projection method and so on.Then,the error analysis commonly seen in the Linear Array 3D SAR imaging is given.The channel error is analyzed and modeled,and the related analysis of the imaging process is given.Through the comparison and analysis of image with the measured data,the correctness of the fixed channel error predicted in advance and the effectiveness of the compensation are proved.Finally,the GPU-based parallelization acceleration method is introduced to reduce the time of the 3D BP imaging algorithm,in order to further shorten the imaging time,we may use the method of randomly extracting the echoes.The noise introduced by the maximum projection imaging and system factors can cause the imaging quality to decline,so we can use some methods like median filtering,image smoothing,image sharpening to effectively improve the quality of the image.3.The foreign object detection based on deep learning technology in security inspection imaging has been completed.Firstly,The important concepts of deep learning,such as non maximum suppression and bounding box regression,are introduced.Secondly,the theoretical basis of the heat map method and the YOLO network are separately introduced.Then,Based on the heat map method and YOLO network,the training is carried out with the measured data of security inspection imaging and the process is analyzed.Finally,through the detection result of two networks with measurement data,we got the comparison of the performance and applicability of two networks in terms of detection accuracy,precision rate and time-consuming.and gave relevant conclusions.4.Traditional SAR image target detection and segmentation experiment based on convolutional neural network is completed.Firstly,each part of convolution neural network is introduced,including convolution layer,activation function,full connection layer and so on.Then,target detection is performed on MSTAR datasets containing background based on YOLO and Fast RCNN respectively.Finally,target segmentation on huge scene SAR image based on full convolution network is realized.