Few-Shot Radar Target Recognition via Meta-Learning and Integrated Time-Frequency Domain Features

Currently, deep neural network is the main method for radar automatic target recognition. Because of the necessity for an extensive volume of annotated data to optimize model parameters, these deep models face serious risks of overfitting in few-short condition. To address the preceding issue, this paper presents a radar few-short target recognition approach via meta-learning mechanism. Firstly, scattering features from geometrical theory of diffraction (GTD) model are extracted to explore more valuable information of the target. Secondly, high resolution range profile data are fused with GTD features by our proposed learner. Thirdly, the proposed meta-learner can use the priori knowledge learned from one polarization state target recognition to guide the learner to better realize another polarization state target recognition, which means the proposed method can perfectly fuse multi-polarization dataset to enhance target recognition performance. Simulation results demonstrate that our method can achieve higher target recognition accuracy compared with conventional deep learning approaches in few-short condition.