24 / 2024-08-03 21:19:55

Advancing Few-shot HRRP Target Recognition with Meta-Learning and Graph Neural Network

HRRP,RATR,Few-Shot Learning,Graph Neural Networks,MAML

Radar High Resolution Range Profiles (HRRP) can provide valuable information for Radar Automatic Target Recognition (RATR). In recent years, HRRP-based RATR methods utilizing deep neural networks have demonstrated promising results due to their remarkable accuracy and performance. However, these deep models are often data-driven, which limits their generalization ability and may lead to severe overfitting issues in few-shot scenarios. This paper introduces a novel few-shot recognition method for HRRP RATR, called Meta-Learning for Graph Neural Networks (MLGNN). This method utilizes the identical ability of few-shot learning Graph Neural Networks (GNNs) to learn the inter-sample relationships between HRRPs, which makes GNNs an efficient approach in few-shot scenarios. In our method, the graph's edges are learned using Multilayer Perceptron (MLP), with HRRP sequences represented as nodes in the graph. Additionally, since the training procedure of few-shot learning GNNs differs from other neural networks, we tailored Model-Agnostic Meta-Learning (MAML) to fit GNNs and finally turn into our MLGNN, a framework for further improve recognition accuracy and generalization performance in few-shot scenarios with few-shot learning GNNs. To prove the efficacy of MLGNN, comparative experiments are conducted on aircraft electromagnetic simulation datasets, whose results demonstrate the superior performance of the proposed method compared to current state-of-the-art meta-learning approaches. Contributing to the remote sensing community, relevant codes is available at: https://github.com/MountainChenCad/MLGNN.