Study Of Radar Target Recognition Based On Supervised Variational Autoencoder

Since High-resolution range profile(HRRP)is easy to store and obtain,it has attracted radar target recognition community's attention.This thesis focuses on the research of target recognition based on HRRP.It proposes convolutional variational autoencoder with learnable prior and convolutional variational autoencoder with disentangled representation.By adding label information into the variational autoencoder model,these models can extract more effective features of high-resolution range profile.The content of this thesis mainly includes the following parts:1.The basic framework of HRRP target recognition is discussed.It mainly introduces the preprocessing step of HRRP,the target feature extraction methods and the classifiers.In the part of the preprocessing step,it mainly focuses on data sensitivity and their corresponding solutions.In the part of target feature extraction methods,it mainly introduces shallow models and deep network models.In addition,it talks about variational autoencoder model(VAE)in detail.Support vector machine(SVM)is talked in detail in the part of the classifiers as well.2.Radar HRRP target recognition method based on convolutional variational autoencoder with learnable prior is discussed.Firstly,convolutional neural network(CNN)is studied,which focuses on convolutional layer,pooling layer,batch normalization layer,fully connected layer and deconvolution layer.Secondly,aiming at solving the shortcomings of the variational autoencoder,convolutional variational autoencoder with learnable prior,a supervised VAE model,is proposed and discussed in detail,which includes the motivation,the loss function and the training step.Finally,experiments are carried out using the measured data.3.Radar HRRP target recognition based on convolutional variational autoencoder with disentangled representation is discussed.This model is also a supervised VAE model.It can not only take the connections between various representations in account,making the representations more reasonable,but also extract more separable features.For the common representation between different classes,the model adopts the representation modeling method similar with VAE.For distinctive representations,the model adopts the representation modeling method similar with convolutional variational autoencoder with learnable prior.By such modelling method,it utilizes the variational lower bound to optimize.After training,the distinctive representations have stronger separability.In addition,the model takes the advantages of convolutional neural networks to extract features.The performance of the model is verified through the measured data.