Stacked Hourglass Network with Integration of Channel Information for Automatic Modulation Recognition

Automatic modulation recognition plays an important role in spectrum monitoring and cognitive radio. However, most of the existing modulation recognition algorithms ignore integration of in-phase quadrature (IQ) channel information, leading to insufficient capability of feature extraction of within and between IQ channels. In order to deal with these defects and improve the recognition accuracy, this paper proposes a new convolutional neural network for automatic modulation recognition, namely stack hourglass network with integration of channel information (HNIC). The network performs better than the baseline hourglass network and other neural networks. It is accomplished due to the following characteristics, such as (1) integration of IQ channel information: to make full use of the channel information of inter and intra IQ channels of the modulation signals, two types of convolution kernels are designed to capture the feature information of the modulation signals. (2) Use of the multi-scale features: the proposed HNIC network obtains multi-scale features via combining the hourglass network with a bottleneck layer, which explores the multi-scale features and increases the depth of the network, improving the accuracy. (3) Network optimization: the structure of the hourglass network transition layer is adjusted, and the channel attention mechanism runs through the up-sampling and down-sampling stages of the network, thus improving the expression ability of the whole network. Experimental results on the benchmark data set RML2016.10a and Augmod show that the proposed method outperforms the existing recognition methods. In addition, the method is of stronger ability of feature extraction, especially for the multi-scale features.