Lightweight Network for Modulation Recognition Based on Stochastic Pruning-Asymmetric Quantization

Automatic modulation recognition (AMR) plays an important role in wireless communication system monitoring, non-cooperative communications, and cognitive communications. Recently, the applications of deep learning in AMR improve classification accuracy. However, it is difficult to deploy a deep learning-based model on resource-constrained devices because of its huge model size. In this paper, we propose a neural network called double pooling convolutional neural network (DP-CNN) and a stochastic pruning-asymmetric quantization (SPAQ) algorithm to realize lightweight and accurate modulation recognition. With the SPAQ algorithm, unimportant parameters are pruned by designing probability intervals and evaluation criteria. In addition, the storage type of parameters will be transformed by creating quantization intervals and mapping criteria. The performance of our method is verified using an open-source dataset RadioML2016.10a. Experimental results show that the SPAQ algorithm has better recognition performance than other lightweight methods at high compression ratios. In addition, the DP-CNN compressed by the SPAQ algorithm outperforms the existing lightweight network in recognition accuracy under the same model size.