DDMNet: A Dual-Stream Network With Differential Temporal and Multiscale Spectral Features for Automatic Modulation Recognition

Automatic Modulation Recognition is a key technology in non-cooperative communications. With the growing complexity of electromagnetic environments, achieving high-precision and low-complexity modulation recognition remains a considerable challenge. Addressing the limitations inherent in existing methods for feature extraction and fusion, this letter introduces a dual-stream network model for modulation classification. This model innovatively incorporates a parallel feature extraction architecture that leverages differential amplitude-phase time-domain features alongside multi-scale spectral features. Furthermore, it achieves deep feature fusion through a dynamic gating mechanism. Experimental results demonstrate that the proposed network achieves an average recognition accuracy of 65.32% and 66.13% on the RML2016.10a and RML2016.10b datasets, respectively, surpassing the state-of-the-art method by 1.53% and 0.66%. This advancement significantly enhances the recognition performance, particularly under low signal-to-noise ratios and for high-order QAM signals.