EUVIMEDEUVIMED entdecken
Crossref

Comparative Study of Time–Frequency Feature Representations for Deep Learning-Based Wideband Signal Detection and Recognition

Xikang Wang, Hua Xu, Zisen Qi, Qingwei Meng, Yixiao Zhou, Yunhao Shi, Wenran Le

IEEE Communications Letters · 2026 · 30 · 2342-2346

DOI: 10.1109/lcomm.2026.3705767 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Communications Letters, 2026. Verfasst von Xikang Wang und 6 weiteren Autor:innen. Die Arbeit nennt 16 Literaturangaben. Gefördert von: National Natural Science Foundation of China.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

As a core technology for intelligent spectrum sensing, wideband signal detection and recognition (WSDR) in the time-frequency domain has attracted extensive research attention in recent years. The mainstream technical paradigm maps the complex-valued time-frequency spectrum to an image representation via logarithmic power transformation, enabling efficient adaptation to well-established general object detection models. However, the transformation from symbolic data to images not only causes phase information loss, but also introduces nonlinear transformations and irreversible mappings, with existing studies lacking effective evaluation of their impact on model accuracy. Therefore, this letter systematically evaluates the impact of ten typical time-frequency feature representations on WSDR performance from energy and phase dimensions. To adapt general object detection models to different time-frequency feature representations, this letter reconstructs the sparse region-based convolutional neural network (Sparse R-CNN) and the detection transformer with improved denoising anchor boxes (DINO) based on real-valued and complex-valued neural networks, respectively. The results verify the superiority of the energy-domain magnitude spectrum for WSDR tasks and demonstrate that both real-valued and complex-valued neural networks exhibit limitations in joint modeling of the magnitude and phase features of the wideband time-frequency spectrum. Our code and dataset are now available as open-source resources at https://github.com/xk-maki/TFR-WSDR

Förderung

  • National Natural Science Foundation of China

Zitieren

APA

Wang, X., Xu, H., Qi, Z., Meng, Q., Zhou, Y., Shi, Y., & Le, W. (2026). Comparative Study of Time–Frequency Feature Representations for Deep Learning-Based Wideband Signal Detection and Recognition. IEEE Communications Letters, 30, 2342-2346. https://doi.org/10.1109/lcomm.2026.3705767

Vancouver

Wang X, Xu H, Qi Z, Meng Q, Zhou Y, Shi Y, et al. Comparative Study of Time–Frequency Feature Representations for Deep Learning-Based Wideband Signal Detection and Recognition. IEEE Communications Letters. 2026;30:2342-2346. doi:10.1109/lcomm.2026.3705767

BibTeX
@article{Wang2026,
  author = {Wang, Xikang and Xu, Hua and Qi, Zisen and Meng, Qingwei and Zhou, Yixiao and Shi, Yunhao and Le, Wenran},
  title = {Comparative Study of Time–Frequency Feature Representations for Deep Learning-Based Wideband Signal Detection and Recognition},
  journal = {IEEE Communications Letters},
  year = {2026},
  volume = {30},
  pages = {2342--2346},
  doi = {10.1109/lcomm.2026.3705767}
}
Kostenlos starten →