Deep Neural Network Compression Technique Towards Efficient Digital Signal Modulation Recognition in Edge Device

Abstract

Digital signal modulation recognition is meaningful for military application and civilian application. In the non-cooperation communication scenario, digital signal modulation recognition will help people identify communication target and have better management over them. In order to the classification accuracy, deep learning is widely used to complete this task. However, current papers have not considered the deployment of deep learning in compute capability and storage limited edge equipment. In this paper, we utilize neural network pruning techniques to reduce the convolution parameters and floating point operations per second (FLOPs), which will pave a wide way to deploy signal classification convolution neural network (CNN) in edge equipment. We set the Average Percentage of Zeros (APoZ) criterion for convolution layers. Compared to original CNN, the experiment result shows that light CNN convolution layer could use only 1.5%~5% parameter and 33%~35% time without losing significant accuracy.