Binary Classification of Modulation Formats in the Presence of Noise through Convolutional Neural Networks

Automatic classification of modulation formats in the presence of noise is a challenging problem which has wide applications in intelligent communication systems. It is a major step towards accurate signal detection on the receiver side given stringent requirements such as spectral efficiency, reliability, latency etc. Deep learning techniques have found wide success in image recognition, detection, speech recognition as well as reinforcement learning domains. In this study, we compared and contrasted 2D and 3D Convolutional Neural Networks (CNNs) for the task of automatic recognition of modulation formats. We used random images of cat, that are modulated using 16 and 64 Quadrature Amplitude Modulation (QAM) formats, passed through Rayleigh and Rician channel formats and classified using both 2D and 3D CNN architectures. We tried 5 and 10 fold cross-validation procedures that are widely used in the literature. We have observed the best performance results in terms of reported metrics of the 3D CNN architecture trained using 10-fold cross-validation and the performance of 2D CNN architecture trained using 10-fold cross-validation procedure to be the worst. We obtained a perfect value of 100% in terms of performance metrics like sensitivity, precision, F-measure, balanced accuracy and specificity, when using 3D architecture trained using 10-fold cross-validation procedure for binary classification among 16 and 64 QAM modulation formats corrupted by Rayleigh and Rician channel noises.