Study Of High-resolution Radar Target Detection And Recognition Based On Auto-encoder

Radar is a kind of all-weather,all-weather radio detection device,which has important practicability in civil and national defense construction such as natural environment detection and battlefield environment perception.With the rapid development of radar sensing technology,it has become a research hotspot to interpret radar data and analyze rich information related to target characteristics.As the key parts in radar sensing field,the radar target detection and recognition have great research value for being the key technologies to interpret radar data.Due to the powerful feature extraction ability of deep neural networks,high-resolution radar target detection and recognition methods based on deep learning have received widespread attention.the existing radar target detection and recognition network relies on a large amount of annotated data to fully learn,and most of them model the amplitude of radar data,which is difficult to use the phase information reflecting the characteristics of the target,and there are problems such as black box and poor network interpretability in the internal process of the network,which greatly limits the performance improvement and promotion of radar target detection and recognition methods.Autoencoder(AE)network is a kind of typical unsupervised neural network framework,which can reconstruct a large amount of unlabeled data to mine useful information and alleviate the problem of poor model generalization caused by limited labels.Besides,AE has good expansibility.On the one hand,AE can be generalized to complex network to make use of phase information in complex radar data to improve the ability of target characteristics description.On the other hand,since the mapping process of decoding network in AE is similar to the process of data generation described by radar physical model,the physical model can be embedded into the decoding network of AE to improve its interpretability.In this dissertation,with two kinds of high-resolution radar data including Synthetic Aperture Radar(SAR)image and High-resolution Range Profile(HRRP),the radar target detection and recognition methods are studied based on the structure of AE for the application scenarios of SAR target detection,HRRP target recognition and SAR target recognition.The main contents of the dissertation are summarized as follows:1.Aiming at the problem that the neural network-based methods rely on a large number of labeled training samples,while the synthetic aperture radar(SAR)images are difficult to be labeled resulting in poor model generalization ability and significant degradation of model performance with limited labels,a domain adaptive auto-encoder network is developed for semi-supervised SAR target detection in Chapter 3.Based on the traditional Faster R-CNN detection model,the decoding and domain adaptation modules are introduced.The decoding module is composed of multi-layer de-convolution networks,which combines the feature extraction network of Faster R-CNN to form an AE branch to reconstruct a large number of unlabeled SAR images,so as to improve the model generalization ability under limited labels.The domain adaptation module transfers the knowledge of abundant labeled optical images to limited labeled SAR image domains and uses optical images to assist the feature learning of SAR image,which can improve the feature separability of SAR images.The experimental results based on the measured SAR image dataset show that our method gains much better detection performance than Faster R-CNN with limited labeled SAR images,and verify the effectiveness of decoding module and domain adaptation module to improve the SAR target detection performance.2.Aiming at the problem that that most current radar HRRP target recognition methods based on deep learning ignore the phase information of radar data,resulting in insufficient ability to describe target characteristics and limited feature separability,Chapter 4 of the dissertation proposes a novel Class Factorized Complex Variational Auto-encoder(CFCVAE)for radar HRRP target recognition.The CFCVAE is a complex AE model that can reconstruct complex HRRPs to make use of phase information to improve the model’s ability of describing target characteristics.To describe observed data precisely,the CFCVAE designs multiple classdecoders in the decoding part of AE by introducing the label information.Each class-decoder describes the generation process of HRRP data belonging to corresponding class,which can improve the feature separability.In addition,the CFCVAE infers the distribution parameters of features by variational inference to improve the generalization of features.In the target recognition stage,based on the minimum reconstruction error criterion,the reconstruction errors of test samples in various class-decoders are compared to determine their categories.The experimental results based on the measured radar HRRP data show that our method can effectively utilize phase information of complex HRRP to improve recognition performance and achieve high recognition accuracy.3.Aiming at the problem that most radar HRRP target recognition networks are of black box structure and are hard to be explained,a complex interpretable probabilistic auto-encoder is proposed for radar HRRP target recognition in Chapter 5.The decoding network of AE combines with the physical generation mechanism of complex HRRP data,and the coding network extracts the target scattering center features,which improves the interpretability of features.In order to accurately describe the generation process of radar HRRPs,our method divides the radar HRRPs into multiple azimuth frames according to the azimuth angle.The complex HRRP in each azimuth frame is reconstructed by using an independent AE,and the number of model’s parameters is reduced by sharing parts of network layers in different AE networks.Due to the physical characteristics that the locations of scattering centers for HRRPs in each frame are similar while their amplitude is fluctuant,our method adopts the way of location sharing and amplitude independence to accurately extract scattering centers.Moreover,the proposed method is a probabilistic model,which uses the probabilistic network to infer the posterior distribution of scattering center features to improve the feature generalization.In the stage of target recognition,our model inputs test samples into different AE networks for reconstruction,and determine the category of samples based on the minimum reconstruction error criterion.The experimental results based on the measured data show that the scattering center features extracted by the proposed method have clear physical meaning and our model achieves high recognition accuracy.4.Aiming at the problem that most of the current SAR target recognition networks have poor interpretability,which leads to poor model generalization for new condition data and rapid recognition performance degradation,Chapter 6 of the dissertation develops a complex autoencoder based on scattering center feature extraction for SAR target recognition.Our method is an end-to-end model combining scattering center feature extraction and target recognition to avoid the problem of mismatch between the features and classifier.Based on the structure of complex auto-encoder,the scattering center model of SAR target is embedded into the decoding network,and physical features containing the locations and amplitude information of target scattering centers are learned by the coding network.Therefore,our model is interpretable and robust to new condition data.Based on the deep learning mechanism,our method avoids the difficulties of determining prior parameters or iteration termination threshold in traditional scattering center extraction methods,and thus can extract scattering centers more accurately.To reduce model’s computational and space complexities,our model adopts the scattering center model in image domain to divide the input images into patches.Since the scattering centers are three-dimensional discrete data containing twodimensional locations and amplitude,our method designs a point cloud network as the classifier to better use the scattering center features for SAR target recognition.Experiments based on measured data show that the proposed method can accurately extract scattering center features for new data under different extended operation conditions(such as the version variant,noise corruption,resolution variant,occlusion and so on),and have achieved robust SAR target recognition performance,with high time-efficiency and low space storage requirement.