IQCM-Net: Cross-Modal Fusion With Multi-Window STFT via SNR Partitioning for Robust Modulation Recognition
Junwei Li, Ming Huang, Jingjing Yang, Zhe Xiao · IEEE Transactions on Cognitive Communications and Networking · 2026
Automatic Modulation Recognition (AMR) is crucial for cognitive radio and spectrum management. However, achieving robust performance, especially under low Signal-to-Noise Ratio (SNR) conditions, remains challenging. This paper introduces IQCM-Net, a novel deep learning architecture specifically designed for enhanced AMR. IQCM-Net leverages two key innovations: a Multi-Window STFT via SNR Partitioning preprocessing strategy, and a Bidirectional Cross-Attention Deformable Convolution (BCADC) fusion module. The Multi-Window STFT approach tailors Time-Frequency (T-F) analysis by assigning pre-optimized window sizes to distinct SNR partitions, improving feature quality across varying noise levels. The BCADC module then effectively fuses temporal and frequency-domain features through dynamic cross-modal attention and receptive fields provided by deformable convolutions. Complemented by a TCN for long-range dependency modeling and a hierarchical feature extraction backbone, IQCM-Net generates highly discriminative representations. Extensive experiments on benchmark datasets (RML2016.10a/b, RML22 and RML2018) demonstrate that IQCM-Net achieves state-of-the-art performance, significantly outperforming existing methods, particularly in low SNR conditions. The results validate the efficacy of the proposed dynamic preprocessing and cross-modal fusion strategies for robust AMR.