Study Of Radar HRRP Target Recognition Based On Probability Statistical Model

An high-resolution range profile(HRRP)denotes the amplitude of the coherent summation of the complex time returns from target scatterers in each range cell,which represents the projection of the complex returned from the target scatting centers onto radar line-of-sight(LOS)and provides valuable information for target classification and recognition,such as target size,target structure,scatterer distribution along the radar LOS,etc.Meanwhile,compared with other wideband signals,e.g.,synthetic aperture radar(SAR)image and inverse synthetic aperture radar(ISAR)image,HRRP has the superiorities of easy acquisition and processing.Therefore,using HRRP to determine the class membership of the unknown target has become a research hotspot in the field of radar automatic target recognition(RATR)community.Due to the solid statistical theory,the probability statistical model,which is built based on the probability theory and mathematical statistics,has unique advantages in radar HRRP target recognition: a)the model can integrate some prior information,which is conducive to target recognition in the case of small samples;b)it can provide the uncertainty assessment of model parameters,effectively alleviating the problem of over fitting and improving the generalization accuracy of recognition;c)it can realize the adaptive adjustment of model structure with data by combining nonparametric Bayesian theory,and so on.Based on the probability statistical model,this dissertation mainly studies the theory and technology of radar HRRP target recognition from two aspects: statistical recognition via the shallow probability statistical models under the condition of small samples and feature extraction via the deep probability statistical models under the condition of large samples.The main contents of this dissertation are summarized as follows:1.Aiming at the bad recognition performance of the factor analysis(FA)model in the case of small training data size,a convolutional factor analysis(CFA)model is proposed.Different from the traditional FA which models the HRRP as a weighted summation of a series of dictionary atoms with the same dimension as the observed samples,the CFA model introduces the convolution operation and represents the HRRP as the summation of convolutions of dictionary atoms and the corresponding weight vectors.As a convolution kernel,each dictionary atom in the CFA model not only has much lower dimension than the observed samples,but also is capable of obtaining the local detail that can better reflect the nature of the data.Compared with the global information,the local information has stronger sharing characteristic.Thus fewer dictionary atoms are required for the description of all observed data via the CFA model.Due to the smaller dictionary atom scale and fewer dictionary atoms,the CFA has lower complexity than the FA model and is more suitable for the small sample target recognition.The experimental results on measured HRRP dataset show that the proposed CFA model has the better recognition performance when there are limited training samples.2.In radar HRRP target recognition based on the CFA model,a statistical model is separately constructed for each frame of each target.Actually,the HRRPs from different frames have certain similarity.Therefore,for all frames of all targets,their corresponding dictionary to generate the observation space should have certain overlaps.Moreover,the CFA model only focuses on the description of observations and does not utilize the supervised information beneficial to the target recognition.According to above analysis,we propose a label constrained convolutional factor analysis model with multi-task leaning,(abbreviated as MTL-LCCFA)on the basis of CFA model.On the one hand,the MTL-LCCFA model share a large dictionary among all data from all targets,and the required dictionary atoms for each observed sample are automatically selected from the dictionary via the non-parametric prior of beta process(BP).Due to parameter sharing,the number of parameters in the convolutional model is further reduced,leading to the less demand for the training samples.On the other hand,the MTL-LCCFA model introduces the class labels of observations to restrain the parameter learning process,which makes the learned model parameters from different classes have greater differences,thus enhancing the separability of the statistical model.Through the parameter sharing and label constraint,the MTL-LCCFA model can further improve the ability in small sample target recognition.The experiments on HRRP dataset show that the statistical modeling with the MTL-LCCFA can strengthen the differences among inter-class models and obtain the better recognition performance with limited training samples.3.Aiming at the fact that the traditional shallow probability model cannot extract the deeper and more effective features of data,a class factorized variational autoencoder(CFVAE)is proposed for the HRRP-based feature extraction and recognition.The CFVAE obtains the probabilistic latent representations of all observed data via an encoding network and constructs a decoding network for each class to describe the generation process from the latent feature distribution of each class to the probability distribution of the corresponding observations.The multi-layer nonlinear structure of the encoding and decoding networks offers the potential to reveal the deeper information of data,which is beneficial to extract the more separable features.Moreover,the modeling approach that a decoding network is built for each class makes each decoder has the better depiction for the data from the corresponding class,and on the contrary,the poor description ability for the data from other classes.Thus the CFVAE can compare the reconstruction errors of a test sample given all the learned decoders to realize the determination of the class membership for the test data,which avoids the mismatch between the extracted feature and the classifier occurred in the traditional variational autoencoder(VAE).Experiments on HRRP dataset show that the proposed CFVAE has the stronger ability to recognize the targets,compared with the shallow models and traditional VAE.4.A discriminative mixture variational autoencoder(DMVAE)is proposed to solve the problem of deep feature extraction under the complicated data distribution and limited labeled training sample conditions.On the one hand,the DMVAE uses the Dirichlet process(DP)to partition the whole dataset into several subsets with the generations of observations in each subset being described via a specific decoding network.The same distributed data is generated by the same decoding network,and different decoding networks represent the generation processes of different distributed data.Since the data distribution of each subset is simpler than that of all observations and each decoding network can more accurately describe the data in the corresponding subset,the collection of multiple decoding networks in the DMVAE offers the potential to give a more precise description to the whole dataset,thus avoiding the poor feature representation ability caused by the inability of a single decoder to accurately describe the data distribution in the case of complex data distribution.On the other hand,the semi-supervised learning mechanism is adopted in the DMVAE,namely,the DMVAE exploits a classification network to represent the generation from the latent features of labeled data to their corresponding class labels while describing the distribution of all observed samples(including labeled and unlabeled samples)via the multiple decoding networks.Through the semi-supervised learning,the DMVAE can not only learn the more discriminative latent space,but also greatly reduces the demand for the labeled training data.Moreover,the classification network restrains the predicted labels of unlabeled samples to have the minimum entropy,which improves the generalization performance of the classifier.Based on the stochastic gradient ascent algorithm,the DMVAE jointly optimizes the marginal log likelihood of all observed data,the label constraint of the labeled samples and the entropy constraint of the unlabeled samples to perform the solution of the model parameters.The experiments based on HRRP dataset validate the advantages of the proposed method in recognition performance,especially in the case of limited labeled samples.