Modulation classification using convolutional neural networks and spatial transformer networks

In this paper, we propose a method for modulation classification based on Spatial Transformer Networks and Convolutional Neural Networks. Spatial Transformer Networks were originally proposed by the computer vision community to make images invariant to spatial transformations. We adopt this model in the modulation classification problem by making the raw baseband I/Q samples invariant for some channel effects. We study the impact of adding STN to CNN classifier in terms of classification accuracy for different oversampling ratios. We also compare the accuracy of proposed classifier with a conventional statistical method based on forth order cumulants. Results show that proposed classifier has better classification accuracy than the one based on cumulants, and its performance is improved for low oversampling ratios.