Automatic Modulation Recognition Based on Asymmetric Convolution and Parameter Fusion

Automatic modulation recognition (AMR) is of great importance in various communications applications. With the rapid development of deep learning (DL), DL-based AMR has been widely used due to its powerful recognition ability. In DL based AMR, both accuracy and complexity are critical performance metrics. However, most existing methods achieve high accuracy at the cost of high complexity. This paper proposes a DL-based AMR method based on the asymmetric convolution and parameter fusion, which improves the recognition accuracy without the increase of model complexity. To improve the recognition accuracy, the asymmetric convolutional network (ACNet) is firstly constructed by replacing all standard convolutional kernels of a baseline model with the asymmetric convolution blocks. To keep the model complexity unchanged, the parameters of ACNet are fused and transferred to a new model, named ACFNet, whose architecture is the same to that of the baseline model. Experimental results verify that the proposed method has higher recognition accuracy but the same model complexity compared with the baseline model.