FFRCNet: Frequency Domain Filtering and Re-Parameterized Convolutional Network for Automatic Modulation Classification

Abstract

Automatic modulation classification (AMC) refers to the identification of the modulation scheme of received signals with little prior knowledge, playing a vital role in ensuring the efficiency and safety of radio communications. However, the high computational complexity limits its application in resource-limited devices. In this letter,we propose a lightweight and efficient frequency domain filtering and re-parameterized convolutional network (FFRCNet) to address this issue. FFRCNet utilizes multiscale convolution and real-valued global filters to extract features and remove redundant information. Additionally, the reparameterization method is utilized to reduce computational complexity and improve the inference speed. Experimental results on the RadioML2016.10a and RML22 datasets demonstrate the effectiveness of the proposed FFRCNet. On RadioML2016.10a, FFRCNet achieves 93.91% accuracy at 6 dB and an average accuracy of 63.63% over all SNRs. On RML22, FFRCNet achieves 93.73% accuracy at 6 dB with an average accuracy of 71.57%. The use of reparameterization method reduces the parameters of FFRCNet by 49.98% and improves the inference time by at least 50.83%.