Multi-shuffled Convolutional Blocks for Low-complex Modulation Recognition

This paper proposes a convolutional neural network architecture for automatic modulation recognition (AMR) in resource-constrained devices, namely MBNet. The architecture comprises a series of connected blocks called shuffled blocks with residual connections, where each block contains grouped convolutional layers and one shuffling module. In the shuffled blocks, grouped convolutions parallelize the convolution process for the low-complex modulation recognition. Additionally, to overcome problems that arise from inefficient group interactions in grouped convolutional layers, a channel shuffling module is deployed to improve the communication among filter groups. Furthermore, MBNet deploys a residual connection from an input of each shuffled block to its corresponding output through an element-wise additional layer to learn spatiotemporal features from in-phase and quadrature signals repetitively to enhance the recognition accuracy. Moreover, to verify the performance of MBNet, experiments were conducted on the RadioML2018.01A dataset. From the experimental results, MBNet achieved high recognition accuracy and trainable parameter utilization efficiency over the state-of-the-art AMR approaches.