GAN-SNR-Shrinkage-Based Network for Modulation Recognition with Small Training Sample Size

Modulation recognition plays an important role in non-cooperative communications. In practice, only a small number of samples can be collected for training purposes. The limited training data degrade the accuracy of the modulation recognition networks. In this paper, we propose a novel network to realize the modulation recognition on basis of the few-shot learning. Generative adversarial networks (GANs) and a signal-to-noise ratio (SNR) augment module are introduced to expand the training dataset. In addition, a preprocessing module and residual shrinkage networks are used to improve the capability of characterizing signal features and the anti-noise performance. The proposed network is evaluated using the RML2016.10a dataset. It is illustrated that the proposed network outperforms the baseline method and the method without data augment with a small number of training samples.