A Novel Modulation Classification Scheme Based on Meta-Transfer Learning with Limited Samples

A novel meta-transfer learning based modulation classification (MTMC) scheme is proposed in this paper, addressing the limitations of deep learning based modulation classification schemes that almost cannot leverage unlabeled data. MTMC effectively leverages unlabeled data and rapidly adapts to new modulation patterns with limited labeled samples. By integrating the strengths of meta-learning and transfer learning, our approach mitigates overfitting issues associated with traditional meta-learning methods for complex networks with small datasets. The MTMC framework comprises three stages: pre-training a network using transfer learning, meta-training by migrating network parameters, and fine-tuning for rapid classification of new tasks. Simulation results demonstrate that the MTMC scheme can significantly enhance feature generalization ability and improve modulation classification accuracy.