Spectrum Enhancement Based Modulation Recognition with Dual-Cue Attention Fusion and Extraction

Automatic modulation recognition (AMR) is a vital stage in spectrum regulation and information security of wire-less communication systems. To explore implicit characteristics of information at low signal-to-noise ratio neglected in previous work, an efficient neural network framework, Dual-Cue Attention Fusion and Extraction Network (DCNet) is designed for AMR tasks. We propose signal distortion compensation by strengthening the effective frequency band features and obtain new semantic features after reconstruction. Additionally, an attention fusion method is developed to integrate semantic cues with signal cues, augmenting overall effectiveness. Ultimately, a multi-scale channel attention operation is executed to extract and amplify multi-scale channel features, thereby enhancing recognition accuracy within the constraints of limited computational resources. Experiments show that the proposed framework outperforms other state-of-the-art (SOTA) models in terms of accuracy with lower computational cost.