Research On Radar One-Dimensional Range Profile Target Recognition With Few Shots

Radar Automatic Target Recognition(RATR)technology is an important branch in the field of target recognition,and has attracted extensive attention from many researchers due to its many advantages.High-Resolution One-Dimensional Range Profile(HRRP)has become one of the main methods for target recognition in the informationized battlefield due to its characteristics of easy acquisition,processing,and fast computation.Datadriven One-dimensional Range Profile target recognition methods have also achieved good results in practical recognition tasks.However,in practical application scenarios,it is difficult to obtain sufficient non cooperative target samples for algorithm training due to factors such as detection conditions and non cooperative nature of the target,which greatly limits the model’s recognition performance.Therefore,HRRP recognition in the case of few-shot has become a key research direction to promote the practical application of technology.This article focuses on the HRRP recognition problem under few-shot conditions and conducts the following research based on deep learning algorithms:(1)To address the problem of limited training samples for some target classes in HRRP,a same-domain few-shot HRRP target recognition method based on statistical feature calibration and dynamic pseudo-label fine-tuning is proposed.For target classes with limited samples,classes with more samples can be used for auxiliary training.The class set with more samples is used as the source domain and the class set with fewer samples is used as the target domain.This method uses statistical features from the source domain to calibrate the distribution of few-shot data in the target domain and uses pseudo-labels generated from the source domain data to fine-tune the model,assisting in training for the target task.This method achieved higher recognition accuracy on four classes of few-shot aerial targets than conventional transfer learning and meta-learning,and can smoothly transition from few-shot tasks to many-shot tasks.(2)To address the problem of limited actual measured samples for all targets,a cross-domain few-shot HRRP recognition method based on feature decoupling is proposed,which achieves transfer from the simulation domain to the actual measurement domain.Considering the class similarity and domain difference between simulation and actual measurement data,Transformer is used as the feature extraction network.The encoder of a variational autoencoder is used as the feature decoupling module to separate domain-independent features and a domain discriminator is used to supervise the learning of the decoupling module,extracting domain-independent features from simulation data to assist in training actual measured data.Multiple experiments were conducted using large amounts of simulation data and small amounts of actual measured data,and the results were analyzed and summarized.The experimental results show that utilizing information from simulation data can significantly improve the effectiveness of actual measured data training,and the proposed method is more effective than directly using simulation data for pre-training.(3)Aiming at the problem of extremely few training samples for some target classes(samples no more than five),a prototype fusion-based method for HRRP recognition with few samples is proposed.This method uses two different learning mechanisms,conventional supervised learning and meta-learning,for parallel training.In the feature space,a prototype estimation is performed for each target class,and a measurement method based on EMD(Earth Mover’s Distance)is used to calculate the similarity between samples.In addition,a prototype fusion method is designed to fuse the prototypes learned by the two mechanisms to obtain more accurate target prototypes,significantly improving the recognition accuracy in the case of very few samples.