Research On Radar HRRP Few-Shot Target Recognition

Radar automatic target recognition(RATR)technology is a hotspot in the field of current target recognition research.The RATR technology is of great significance for improving the national defense capability and the level of military command automation.In recent years,with the rapid development of wideband radar technology,the resolution of radar signals has been continuously improved,and the research and application of radar target recognition based on HRRP has emerged.But when the HRRP dimension is too large relative to the number of samples or tracking non-cooperative targets,few-shot problems will occur.This thesis mainly studies RATR method based on HRRP and the few-shot target recognition problem.The main contents and results are as follows:(1)Based on the scattering point model of high resolution range profile,this thesis introduces the basic characteristics and preprocessing methods of high resolution range profile,and then introduces commonly used recognition methods,including adaptive Gaussian classifier,support vector machine and so on.Using the measured data and the above methods to carry out the target recognition experiment,and analyze and summarize the experimental results,compare the performance differences of various methods,analyze the advantages and disadvantages of various methods,and give the opinion of selecting the radar target recognition method.(2)In the few-shot scenario,the direct training classifier will have over-fitting phenomenon.The metric learning does not directly classify,but learns a distance metric function,which can describe the distance(or similarity)between two samples..This kind of thinking is very suitable for solving few-shot recognition problems.The rapid development of neural networks in recent years has made people realize their powerful capabilities.Many researchers have combined neural networks with metric learning and used neural networks to learn distance measurement functions,and many research results have been obtained.This thesis studies the few-shot target recognition methods based on metric learning,including siamese neural networks and prototypical networks.Both methods map the original sample to the feature space through the neural network,calculate the distance in the feature space and complete the few-shot recognition task.Finally,the effectiveness of the two methods over traditional methods is verified by experiments,and the advantages and disadvantages of the two methods are analyzed.(3)Meta-learning extends the learning range of the learner and to learn on the distribution of related tasks.Traditional learner learns the general rules that apply to individual sample points,whereas meta-learner learns an algorithm that works for a variety of tasks.This thesis adopts a meta-learning method to solve the problem of radar few-shot target recognition,namely Model-Agnostic Meta-Learning(MAML).The key idea underlying MAML is to train the model's initial parameters such that the model has maximal performance on a new task after the parameters have been updated through one or more gradient steps computed with a small amount of data from that new task.Applying MAML to the radar high resolution range profile few-shot recognition task,it is verified by experiments that the MAML method can adapt to the new task quickly and obtain better recognition performance.