Few-shot Radar Target Recognition Based On Meta-learning

Radar automatic target recognition(RATR)is a hot topic in the field of radar.RATR plays an important role in both military and civilian fields.RATR methods based on deep learning have shown progressive performance thanks to their advantage of automatic feature extraction and strong generalization ability.However,most RATR methods based on deep learning need a lot of labeled data for parameter optimization,otherwise they world possibly encounter serious overfitting problem,which leads to the low recognition accuracy and poor generalization ability.In order to solve above problems,we studied on the research of few-shot radar target recognition technology and proposed three few-shot radar target recognition methods,which improves the recognition accuracy and generalization ability.The main contents of this paper are summarized as follows:1.A novel few-shot HRRP target recognition method based on meta-learning framework is proposed,which introduces Long Short-Term Memory(LSTM)based neural network as learner for statistical HRRP data.The proposed method exploits multi-polarization HRRP data for RATR and successfully improves recognition accuracy and generalization performance in few-shot condition.We evaluated our method based on an electromagnetic calculation dataset of airplanes and found that the proposed method could successfully fuse multi-polarization HRRP data to provide more effective information for RATR.The experimental results also showed that the proposed method produced improved performance compared with state-of-the-art fewshot learning methods.2.A novel few-shot SAR target recognition method based on gated multi-scale matching network is proposed,which introduces weight gated unit and multi-scale feature extraction module into matching network.In the proposed method,the multiscale feature extraction module is used to extract multi-scale features of different convolutional layers in matching network and the weight gated unit is used to weight different multi-scale features according to different recognition tasks.The proposed method achieves the effect of carrying out different recognition tasks mainly based on features of different layers thanks to the weight gated unit,which can weight different multi-scale features according to different recognition tasks.The proposed method is evaluated on the public measured dataset MSTAR and achieved promising performance compared with state-of-the-art few-shot learning methods and few-shot SAR target recognition methods.Furthermore,the proposed method shows good robustness in noisy environments.3.A novel few-shot SAR target recognition method named Meta-Res Net is proposed.In this method,we designed a novel learner based on residual network,which can effectively transmit contrast information in SAR images and thus improve recognition accuracy.The improved meta-learner can not only learn good initialization parameters for learner,but also learn a different but appropriate learning rate for each learner parameter.The comparative experiment between the proposed method and other three few-shot recognition methods demonstrated the effectiveness and progressiveness of the proposed method.We also conducted experiment to verify the robustness and study the influence of the network structure on recognition accuracy.We showed the different but appropriate learning rates of learner parameters learned by the meta-learner.