Phase-Aware Signal Detector for Accurate Classification and Time-Frequency Localization in Wideband Spectrogram

The wideband signal detection framework which applies deep learning-based object detection networks to wideband spectrograms for joint signal detection, classification, and time-frequency localization has attracted growing interest. However, it remains an open question whether this approach can effectively distinguish PSK/QAM modulation signals with visually confusing time-frequency characteristics in spectrograms. This paper proposes a phase-aware spectrogram (PA spectrogram) to answer this question positively. Furthermore, previous state-of-the-art networks designed for wideband signal detection mainly employ anchor-based and direct localization regression-based detection paradigms. Nevertheless, the diversity of signal transmission parameters results in significantly varying scales and aspect ratios of signal bounding boxes in spectrograms as well as large regression variance for localization. These factors result in the anchor mismatch problem and learning difficulty of direct regression, ultimately leading to suboptimal detection performance and inaccurate time-frequency localization. To address these problems, we propose a novel detection network based on the dual-granularity cooperative localization, termed DGCL-Net, which employs a concise anchor-free paradigm and a coarse-grained classification plus fine-grained regression strategy to achieve more accurate time-frequency localization. Experiments on synthetic and real signal datasets demonstrate the effectiveness of the PA spectrogram and the superiority of the DGCL-Net compared to other baseline networks.