SAR Image Target Detection And Recongnition Based On Deep Learning

Synthetic aperture radar(SAR)has become an important means of information acquisition in military and civilian fields due to its unique advantages of omnidirectional,all-weather and all-day monitoring capability.As the SAR acquisition technology becomes more and more mature,it brings new challenges to the interpretation of SAR images.As one of the key technologies of SAR image interpretation system,target detection and recognition technology has always been a research hotspot in various countries.In recent years,the rapid improvement of computers computation and performance makes the acquisition of big data become possible,and promotes further the development of deep learning,especially in target detection,image recognition fields.However,the traditional SAR image target detection and recognition method only explores the statistical information of the SAR image,and lacks sufficiently mining of the SAR target features,which results in poor performance of the target detection and recognition.Therefore,in this thesis,for the problems of traditional SAR image target detection and recognition methods,the deep learning with the excellent feature extraction ability is adopted to exploit the SAR target features to improve the detection and recognition performance.Meanwhile,the problems of massive SAR image interpretation can be alleviated to some extent.This paper mainly conducts research on SAR image target detection and recognition based on deep learning theory.The main contents and innovations are as follows:(1)For the problem that the traditional convolutional neural network(CNN)structure has the low SAR image classification performance due to ignoring the important low-level feature information,in this paper,we propose a SAR image target classification method based on multi-feature combination.Firstly,the method visualizes the convolutional feature map of ZF-Net network and analyzes the limitations of exiting training network in SAR image target classification.And then,a novel feature extraction method of low-level texture features and high-level semantic information is proposed,and the extracted features are input into the full connection layer to classify the SAR image target.In addition,for the overfitting problem caused by the insufficient samples of MSTAR data set,a sample data enhancement method based on Markov random field is designed.Finally,the validity of the algorithm is verified by the measured data.(2)For the problem that the traditional SAR image target detection and recognition algorithm process is a hierarchical attention mechanism,which leads to cumbersome detection and recognition process and large amount of calculation,in this paper,we propose an end-to-end SAR image target detection and recognition method.Firstly,this method utilizes the region proposal network(RPN)to extract the SAR target candidate regions,and the candidate regions are classified and identified by the CNN based on multi-feature cooperation proposed in this paper.Compared with the mainstream ZF-net,VggNet and ResNet,the classification and recognition performances are analyzed.In addition,for the MSTAR data set with the overfitting problem due to the single target and simple scene,in this paper,SSDD data set is adopted to realize SAR multi-target detection and recognition in complex scenes,and verifies the effectiveness of the proposed method.