Study Of Radar High Resolution Range Profile Target Recognition Based On Convolutional Neural Network

Wide-band radar can obtain high resolution range profile(HRRP)of the observation target,HRRP reflects the distribution of the target scattering centers along the line of sight of the radar.HRRP consists of abundant target feature information,and has the advantages of small calculation burden and easy data acquisition way.It is widely used in radar automatic target recognition(RATR).In recent years,deep learning technology has developed rapidly and is widely used in computer vision and other fields.As a typical deep learning model,convolutional neural network(CNN)has achieved huge success in image recognition,natural language processing and other fields.Therefore,it is of great significance to develop HRRP target recognition method based on CNN.Hence,the study of this thesis can provide important theoretical value for improving RATR technology of our country.The main contents of the thesis are presented as follows:(1)The basic principles of wide-band radar range imaging and typical target recognition methods are introduced.Firstly,based on the classic turntable model,the basic principles of inverse synthetic aperture radar(ISAR)one-dimensional range imaging are introduced.Then,the influences of the amplitude,translation and attitude sensitivity characteristics of the HRRP on recognition performane are analyzed.The corresponding preprocessing methods are given.Finally,HRRP target recognition based on support vector machine(SVM)and k nearest neighbor(KNN)algorithm are introduced to provide a basis for the performance analysis and comparison in subsequent chapters.(2)Study on CNN-based HRRP target recognition method.Firstly,the structural characteristics of the CNN,the basic principles of network training and optimization algorithms are introduced.Then,by analyzing the characteristics of HRRP,a HRRP target recognition method based on the typical CNN is designed.The network can extract translation invariant features through pooling structure,which effectively eases the translation sensitivity problem of HRRP,without the preprocessing of envelope alignment.Hence,the "end-to-end" output from HRRP data to recognition results is realized.Finally,based on the measured HRRP data of Yak-42,An-26 and Citation aircrafts,support vector machine,k nearest neighbor and CNN are used for target recognition.The results show that the CNN has great advantages in recognition accuracy and noise robustness.(3)A batch normalization and feature fusion-based CNN is proposed to achieve the HRRP target recognition.Firstly,the limitations of traditional CNN structure in feature extraction is analyzed.Then,the basic principles of the feature fusion network and batch normalization are introduced respectively.The feature fusion network can merges the features of different layers,e.g.the structure of the shallow network and the details of the deep network.The batch normalization(BN)operation can speed up the network training efficiency and the convergence,suppress the over-fitting,and enhance the network generalization capabilities.Then,a new BN and feature fusion-based CNN structures are proposed to achieve the HRRP target recognition.Finally,experiments on the measured HRRP data sets of three classes of aircrafts show that the method proposed in this chapter has better performance than the typical CNN structure.(4)Conclusion and the prospect of the future work.The work of this thesis is concluded and the future work is prospected.