Auto Encoder Based Methods For High Resolution Radar Target Recognition

High-resolution radar data is widely used in radar automatic target recognition(RATR)because it can provide target size,scatterer distribution and other characteristics,and contains detailed information of the target.Thus,high-resolution radar data can distinguish the type of target precisely.With the rapid development of deep learning in the field of machine learning,deep learning methods are also introduced into the application of RATR and encourage the further development of RATR technology.Among deep learning methods,AE is widely used in various areas due to its good generalization performance.This dissertation studies the theories and techniques of deep learning and focuses on problems when applying AE to high-resolution radar data for RATR.Our researches are supported by Advanced Defense Research Programs of China and National Science Foundation of China.The main works of this dissertation is summarized as follows.1.We firstly introduce the basic concepts and the fundamental theories of RATR,and review several related works in recent years.The basic concepts and recent works of deep learning are introduced in the second.Finally,a brief description of the main works in our dissertation is given.2.We introduce the basic structure of restricted Boltzmann machine(RBM)and autoencoder(AE),which are typically used deep learning methods.The detailed derivation of RBM and AE are explained and then the characteristics and differences of the two methods are analyzed.Furthermore,the performance of the two methods is analyzed by the experimental results on HRRP data and SAR images in RATR.3.We study the problem of directly applying AE to high-resolution radar data for RATR.Lack of training samples may lead to the decrease of recognition accuracy.Data augmentation methods,which are designed to deal with small amount of training data,are hard to use in real applications,because which manner of augmentation yields the highest accuracies are usually unknown.To solve the problem,in this chapter we propose the Euclidean distance restricted AE,which introduces the supervised constraint of Euclidean distance to the original AE.By using the supervised information of training data,the recognition performance can be improved without data augmentation.Furthermore,the dropout step is added to the Euclidean distance restricted AE to avoid over-fitting caused by the small amount of training data and the supervised constraint.Finally,the effectiveness of the Euclidean distance restricted AE is proved by the recognition results on SAR images,and the performance of our proposed method is analyzed.4.We study the problem that training samples are usually polluted by noise and clutter background in real applications of RATR on high resolution radar data.Most of the deep algorithms ignore to consider the target area and the background area separately in feature learning.Thus,these methods learn hierarchical features from both target area and the noise or clutter background area and the performance are influenced by the noise or clutter background.To solve the problem,in this chapter we propose the point-wise discriminative AE(PDAE),which adopted a specially designed target area extraction net to separate the target area from the noise or clutter background.Furthermore,the supervised constraint is introduced into PDAE,which makes the target area extraction network extract the target area precisely and further extract the discriminative high-level features from the target area for recognition.Point-wise gated Boltzmann machine(PGBM)is also introduced in this chapter for comparison.Finally,the effectiveness of PDAE is proved on HRRP samples of different signal to noise ratios(SNR)and SAR images of different signal to clutter ratios(SCR),and the performance of the target area extraction net is analyzed.5.In the application of high-resolution radar target recognition,the number of training samples may be small,and the samples may be polluted by the noise or clutter,so the performance of directly extracting features by deep learning methods are not good.In real applications,due to the accumulation of experience,many manual features have been extracted for recognition.However,due to the limitation of human experience,those manual features are not all suitable for target recognition.Therefore,to solve the problem of how to utilize the manual extracted features more effectively by deep learning methods,in this chapter we propose the feature selection AE.The proposed introduces the supervised constraint to filter those manual features and further extracts discriminative high-level features by multi-layer nonlinear mapping.Traditional feature selection methods are introduced in this chapter for comparison.Three translation invariant features for HRRP recognition and SIFT features for SAR image recognition are also introduced in this chapter.Finally,the effectiveness of feature selection AE is proved on translation invariant features of HRRP data and SIFT features of SAR images,and the feature selection performance of the proposed method is analyzed.