Adaptive Denoising With Efficient Channel Attention for Automatic Modulation Recognition

Automatic modulation recognition (AMR) is crucial for efficient modulation type recognition in modern wireless systems, but its performance suffers from low signal-to-noise ratios (SNR). This paper proposes an adaptive denoising automatic modulation recognition network (AD-AMR Net) to improve its performance, which incorporates an adaptive denoising module (ADM) to alleviate noise and a feature extraction module (FEM) to extract multi-scale features from the denoised signals. Com-pared to conventional AMR schemes, AD-AMR Net demonstrates better recognition accuracy under low SNRs, achieving 84.8% average accuracy in the 0 – 10 dB range versus 80 – 82% for conventional methods. Overall, the proposed model achieves 64.6% accuracy on the RadioML dataset, substantially outperforming previous deep learning techniques. Key innovations of adaptive denoising approach and multi-scale feature extraction enable major performance gains in low SNR scenarios.