IQTF-GFN: A Gated dual-path network for automatic modulation classification of dual-component overlapping emitter signals

Achieving reliable modulation recognition for overlapping dual-component emitter signals in dense, low signal-to-noise ratio (SNR) electromagnetic environments presents a formidable challenge for modern electronic reconnaissance. Conventional deep learning approaches, which primarily rely on time-frequency images (TFIs), often overlook intrinsic physical properties of signals, resulting in substantial performance degradation under severe noise conditions. To address this, we propose the I/Q and Time-Frequency Gated Fusion Network (IQTF-GFN), a novel cross-modal fusion framework that systematically integrates physical priors into advanced deep learning architectures. The framework employs a parallel dual-branch structure to jointly process one-dimensional (1D) I/Q sequences and two-dimensional (2D) TFIs. Its key innovations include the incorporation of physical priors into the I/Q branch via a Higher-Order Statistics (HOS) pathway for noise-invariant feature extraction, an attention-driven Multi-Instance Learning (MIL) mechanism in the TFI branch to adaptively emphasize salient spectral regions, and a task-aware gating network for dynamic and intelligent fusion. Extensive experiments across 10 Monte Carlo trials demonstrate that IQTF-GFN sets a new state-of-the-art (SOTA) benchmark in both robustness and efficiency. Under the challenging condition of −9 dB SNR, the framework achieves an average Exact Match Ratio (EMR) of 94.11%, outperforming the strongest baseline by more than 11 percentage points. Remarkably, this performance is delivered by a highly efficient architecture with only 19.64 million parameters and 1.38 GFLOPs. The design reduces the theoretical computational load by up to 78% and achieves a practical inference latency of just 0.52 ms per sample. By combining high accuracy with computational efficiency, IQTF-GFN provides a robust and practical solution, introducing a new paradigm for embedding physical priors into deep learning for complex electromagnetic signal recognition.