Feature Learning Methods Based On Deep Generative Networks

With the development of society,science and technology,and the deepening of information technology,data containing a lot of information is constantly produced,but at the same time there are more and more redundant information.Research on how to quickly extract valuable information from massive data and apply it to different tasks is of great significance for national defense and people's livelihood.However,the data in practical application is not only diverse in form,complex in structure and strong in variability,but also particularly ”large”,that is,large both in quantity and dimension.In this case,the problem of efficient feature learning for such large data is extremely challenging.Recently,deep learning has performed amazingly in many tasks related to people's lives and even surpassed human's performance.For a large number of high-dimensional data,deep models can hierarchically learn the shared features of the data and represent the data features in different layers.It is precisely the hierarchical feature representation that makes deep models very suitable for dealing with such data with a large number of dimensions.This paper focus on feature learning tasks in the case of large data,and apply the high-performance model of deep learning to various tasks in the field of national defense and life combined with their characteristics.The main contents of the paper are summarized as follows:1.Radar target recognition technology plays an irreplaceable role in improving the combat effectiveness of our army and stabilizing national defense.The problem of radar target recognition based on HRRP data has attracted extensive attention of researchers in this field because of its effectiveness and convenience.In this paper,we propose a conditional generation model for radar high resolution range profile(HRRP)target recognition.MLPs are used as the sufficient statistics of posterior approximation distribution to learn recognition features and fully encode the variability of observed data features,thus providing potential for improving overall recognition performance for radar high resolution range profile(HRRP)target recognition.Considering the target azimuth sensitivity of HRRP,the model is regularized by reconstructing the average range profile.Then,a three-way weight tensor is introduced to capture the multiplicative interaction between label information and HRRP samples,and further decomposed to effectively reduce model parameters.A lot of experiments on the measured HRRP data show that the algorithm has a good performance of target recognition and reconstruction.2.Considering the sensitivity of target translation in radar target recognition task based on HRRP data,this paper discusses the method of introducing hidden random variables into the hidden state of recurrent neural network(RNN)combined with the basic principle of VAE.We believe that TVAE can better model the variability observed in highly structured sequence data such as HRRP by using more flexible implicit random variables in hidden state.In order to introduce the label information,we use conditional prior theory to make the learned hidden space conditioned on the labels of samples.This develops the model to be a supervised one,which is suitable for the recognition task of HRRP sequential data after interception.In order to further reduce the model parameters and accelerate the calculation speed,we factorize the weights of the model.We prove the temporal prediction ability of the model on the video dataset and the recognition ability of the model on the measured HRRP data.3.Compressed Sensing(CS)theory is one of the effective methods to solve the problem that the amount of data to be transmitted,stored and processed is too huge.The object of compressed sensing is to estimate the original vector from the under-determined system of linear measurement with noise by using the prior knowledge of vector structure in the data correlation domain.Some existing methods,such as CS algorithm for generative model in specific domain,can reduce the requirement of signal structure's sparseness and the number of restored measurements,unlike the traditional CS algorithm based on sparse prior.But they can only ensure accurate recovery by searching for the optimal input based on a twostage strategy.We propose a method of joint training inference model,generative model and discriminative model based on VAE and GAN.In this way,the information in measurements will affect the process of feature extraction and original image generation,which is ignored in previous models.In addition,with the help of inference networks,our model can process test data through fast mapping rather than iterative optimization.In order to make full use of the information retained in the measurements,we also propose a deep conditional generation model to achieve better structured output to restore the original images.Compared with the existing CS recovery algorithm based on generative model,we have made significant performance improvements in reconstruction error and time complexity under different measurement rates on the standard image datasets.Compared with the traditional CS methods,our model has a great improvement in performance under the condition of very few measurements.