Radar HRRP Recognition With CNN And Feature Fusion Methods

Automatic Target Recognition(ATR)plays an important role in the modern information wars.With the development of radar imaging and information processing technology,the target recognition based on radar High Range Resolution Profile(HRRP)has become a hot research topic.In order to avoid the difficulties of extracting good features manually and solve the shift sensitivity of HRRP,the HRRP target recognition based on Convolutional Neural Network(CNN)and an adaptive weighted fusion algorithm is investigated in this thesis.First of all,the basic knowledge related to HRRP and CNN is introduced,which includes HRRP profiling,target recognition methods,the origin and development of CNN,the network structure of CNN and Caffe deep learning library.Then,one-dimensional CNN is applied to HRRP recognition to avoid the difficulties of extracting good features manually.The CNN is trained to extract useful features from HRRP hierarchically and classify targets automatically.The configuration of CNN is described,and the recognition performance of different activation functions,parameters,network structures and different classifiers is compared by simulations.Visual feature maps demonstrate that CNN can obtain distinguishable features through convolutional layer.Aiming to the problem of shift sensitivity of HRRP,an adaptive weighted fusion algorithm is presented.Three one-dimensional frequency domain shift invariant features are weighted,and the weighting coefficients are optimized by BP algorithm dynamically.In the simulation experiments,the recognition performance of FFT,power spectrum,differential power spectrum and different fusion methods is compared.The experimental results show that CNN can learn useful features effictively from HRRP and have a good recognition performance.The adaptive weighted feature fusion algorithm can effectively optimize the weighting coefficients,and also achieved good recognition performance.