A Lightweight Modulation Recognition Network with Depthwise Separable Convolution and SNR Enhancement

This paper addresses the lightweight design requirements for edge device deployment and proposes a lightweight modulation recognition network (SNR-DSNet) based on depthwise separable convolution and SNR feature calibration. The model employs depthwise separable convolutions, reducing FLOPs by 72.73% versus standard convolutions. An SNR-conditional calibration module maps SNR values to 128 - dimensional feature weights via an MLP, dynamically adjusting features to improve robustness across signal quality conditions. On the RML2016.10a dataset, the model with 116k parameters achieves an average test accuracy of 70.29%, indicating its superiority in balancing lightweight design and recognition performance.