Automatic Modulation Classification using Relation Network with Denoising Autoencoder

Abstract

With the advent of deep learning (DL), various automatic modulation classification (AMC) methods using deep learning architectures achieved significant performance improvements compared to conventional algorithms. Aiming to achieve high classification accuracy, DL-based AMC algorithms require numerous annotated training samples for each modulation class to extract salient features, but it is hardly applicable in real-world AMC applications. To tackle the annotated data scarce issue, this paper proposes a novel few-shot learning (FSL) framework, which introduces a relation network with a denoising autoencoder to extract feature representations effectively from a limited dataset. The experimental result demonstrate that the proposed method can achieve higher classification accuracy compared to the conventional FSL algorithm for signal modulation recognition, especially in low signal to noise ratio conditions.