This is the authors' abstract. We don't add key points for this paper.
Automatic Modulation Recognition (AMR) plays a crucial role in cognitive radio and complex electromagnetic environments. However, the efficacy of existing AMR techniques often declines significantly in low-to-intermediate Signal-to-Noise Ratio (SNR) scenarios due to signal distortion. To address this, this paper proposes a novel framework, the Fully Spatio-Temporal Dense Neural Network (FSTDNN). Unlike passive classifiers, FSTDNN first integrates a tailored Generative Adversarial Network (GAN) for active signal restoration, which maps corrupted observations to clean manifolds and recovers spectral fidelity (correlation > 0.99). Following restoration, a unified backbone combining Densely Connected Networks (DenseNet) and Bidirectional Long Short-Term Memory (BiLSTM) is employed. Furthermore, to ensure training stability and deep feature fusion, FSTDNN constructs dual-level gradient pathways via skip connections both within dense blocks and across stacked modules. An embedded attention mechanism further refines the network by prioritizing essential feature channels. Experimental results demonstrate FSTDNN’s superiority: it achieves peak recognition accuracies of 93.08% on RML2016.10a and approximately 94% on RML2016.10b. Notably, in challenging low-to-intermediate SNR regimes (−10 dB to 0 dB), it outperforms the mainstream algorithms by 2% to 7%.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Luo et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: