Sparsely Connected CNN for Efficient Automatic Modulation Recognition

This paper proposes a convolutional neural network (CNN), called SCGNet, for low-complexity and robust modulation recognition in intelligent communication receivers. Principally, the network combines two types of sparse convolutional layers-depthwise and regular grouped in an architecture to achieve high recognition accuracy while keeping the network more lightweight. The network architecture leverages sparsely connected convolutional layers in three principal modules: speed-accuracy tradeoff (SAT), deep feature extraction and processing (DFEP), and generic feature extraction (GFE) data pre-processing module. For a good tradeoff between complexity and accuracy, SAT deploys depthwise convolutional layers to enrich the relevant features outputted by the former GFE module. In addition to SAT, DFEP employs a cascade of regular grouped convolutional layers for mining more discriminative features from SAT via a multilayer transformation module. This cascade structure aims to prevent a loss of essential details of the signal as the network becomes deeper. Additionally, skip connections are deployed between sub-blocks within SAT and DFEP to allow inter-module feature sharing and to handle inter-block features loss. Experimental results on the RadioML2018.01A dataset indicate that SCGNet achieves an overall recognition accuracy of around 94.39% at a signal-to-noise ratio of +20 dB.