A Study On Methods Of SAR Image Target Automatic Detection And Recognition

With the increasing number of Synthetic Aperture Radar(SAR)satellites,the number of SAR images obtained is also growing in a geometric series.However,the ability to interpret SAR images is lagging behind.The target detection and recognition of SAR image is the main target of SAR image interpretation.At present,the target detection and recognition of SAR image is mainly interpreted by artificial interpretation,the low degree of automation can not meet the needs of engineering application.How to improve the automation of target detection and recognition of SAR images and improve the efficiency and accuracy of detection and recognition has become a research hotspot in recent years.SAR image target detection and recognition include SAR image denoising,SAR image segmentation,Target detection and Target recognition.The efficiency and accuracy of SAR image target detection and recognition can be finally improved only by optimize that algorithms used in these processing step.Therefore,It is very important to study the method of each step in the process of sar image target automatic detection and recognition for promoting the wider engineering application of SAR.This paper mainly studies the SAR image characteristics,image denoising,image target segmentation,SAR image target detection based on Beamlet transformation and SAR target recognition method based on deep learning.Based on the analysis of the principles and characteristics of SAR imaging,the paper first need to preprocess the image denoising and target segmentation,which is the work to be done for target detection and recognition.Therefore,the characteristics of SAR image are studied and analyzed firstly in the paper,aiming at the speckle noise caused by the SAR coherent imaging mechanism,based on the analysis of the existing SAR image denoising methods,the Bandelet transform in the hyper-wavelet transform is applied to the SAR image preprocessing engineering practice,Bandelet transform can retain the direction and edge information of the image more in image denoising,and it has some advanced features.Secondly,in the SAR image segmentation,a SAR image segmentation method combined with the global Maxflow Otsu segmentation algorithm and the neighborhood growth algorithm is proposed.The method is robust to SAR image segmentation,and the result is good.The threshold threshold is selected to filter the interference in the image and improve the target recognition rate.Thirdly,on the basis of the research and analysis of the existing target detection methods,aiming at linear objects,the theory of Beamlet transformation is studied,and the Beamlet algorithm is applied to SAR image target detection.A SAR image target detection method based on Beamlet transform is proposed,using Beamlet transform image contour description and encoding,quadtree Beamlet of multiresolution analysis,strong noise background in the target line extraction,distribution of galaxies and target shape progressive encoding,and applied to the video sequence encoding the target shape,and achieved good effect,the method overcome the influence of background noise on target detection extraction.In the last,facing the present situation of the classification recognition accuracy of the established SAR automatic target recognition system is not high enough,based on the research and analysis of the theory and development of convolution neural network,a series of new models,such as LR,LR-1 and LR-2,are modified by modifying the mainstream framework LeNet-5 model in the field of target recognition,by testing 3 kinds of maneuvering targets,such as BMP truck,BTR armored vehicle and T72 tank,the classification accuracy of the target is 94.5%,98.6% and 99.6% respectively.The main innovations in this paper are as follows:1.A SAR image segmentation method combined with global Maxflow Otsu segmentation algorithm and neighborhood growth algorithm is proposed.Using the two kinds of information of global Maxflow algorithm,texture and boundary,to segment the image,creatively improve the field growth method based on morphology,effectively suppress the interference,and lay the foundation for improving the target recognition rate.2.A target detection method for SAR image based on Beamlet Transform is proposed.Compared with the traditional edge detection operators(such as Sobel operator,Robert gradient operator,Prewitt operator,Log operator,Canny operator,etc.),this method can effectively eliminate the influence of image noise,detect the edge of target better,and improve the accuracy of target detection.3.A new LR series model based on LeNet-5 model is proposed.According to the situation of the constructed SAR automatic target recognition system of classification accuracy,based on analyzing the theory of convolutional neural network and its development,through continuous improvement of target recognition in the field of the mainstream framework of the LeNet-5 model,LR,LR-1,LR-2 series model,feature extraction method compared with the traditional SAR image recognition method in(such as the Mellin transform,time-frequency analysis,wavelet transform,neural network,high order moment,SVM etc.),the model can effectively enhance the classification accuracy,higher degree of automation.The process of SAR image target detection and recognition as the main line of the paper,based on the characteristics of SAR image deeply,aiming at the existing deficiencies of the processing method of SAR image denoising,image segmentation,target detection and target recognition and other key steps,some improved methods are presented respectively.The purpose is to efficiency and precision to enhance SAR image target detection and classification,improve the degree of automation of SAR image target detection and recognition,in order to better meet the needs of engineering applications,so as to promote the image of SAR in military and civil applications.