A Denoising Radio Classifier with Residual Learning for Modulation Recognition

Deep learning has great potential in modulation recognition. However, when noise causes signals distortion, the performance of deep learning algorithms degrades dramatically, and the performance deterioration is especially noticeable at low SNR. In this paper, a deep learning-based classification architecture is proposed to enhance modulated signals. The architecture consists of a denoiser and a classifier. The denoiser based on generative adversarial networks (GAN) with residual learning, which implicitly removes the potentially clean signals from the original modulated signals to learn the distribution of noise, and finally enhances the original modulated signals. The classifier is the radio classify convolutional neural network (CNN) based on the CNN. Compared with some renowned machine learning methods and deep learning methods used in modulation recognition, this method can obtain higher classification accuracy, especially when the SNR is low.