FAE-MSLKNet: Multi-Scale Large Kernel Network With Fourier Adaptive Enhancement for Automatic Modulation Recognition

Automatic Modulation Recognition (AMR) is vital in wireless communication systems. Recently, a number of deep learning (DL) architectures have been developed for AMR. However, existing methods have limitations in frequency analysis and long-range temporal dependencies extraction. Moreover, most models have high computational costs and large model sizes. To overcome these limitations, this paper proposes FAE-MSLKNet, a Multi-Scale Large Kernel Network with Fourier Adaptive Enhancement. First, we introduced a Fourier Adaptive Enhancement (FAE) module to adaptively enhance features and model long-term temporal dependencies in the frequency-domain. Furthermore, a Multi-Scale Large Kernel (MSLK) module is employed to extract local features with small convolution kernels and capture long-range temporal dependencies with large convolution kernels in the time-domain. Experiments on RML2016.10a, RML2016.10b, and RML2018.01a datasets demonstrate that FAE-MSLKNet achieves state-of-the-art (SOTA) performance with improved parameter efficiency and reduced computational complexity, highlighting its potential for practical wireless communication applications.