Scholarlink

<p>Traditional modulation recognition struggles with multi-signal coexistence and low signal-to-noise ratios. This paper proposes an integrated framework for wideband signal detection, recognition, and parameter extraction using object detection technology. By converting signals into spectrograms via Short-Time Fourier Transform, the framework employs RT-DETR as the backbone for feature extraction. Key modules, including Transformer-based Intra-scale Feature Interaction and CNN-based Cross-scale Feature Fusion, are introduced to overcome local convolution limitations, enabling precise end-to-end localization and classification. Furthermore, the system directly extracts physical parameters, such as center frequency and bandwidth, from predicted bounding box geometries. Using a dataset of nine modulation types, experiments across an SNR range of − 30 dB to 30 dB demonstrate that RT-DETR outperforms You Only Look Once (YOLO)v8, YOLOv10, and YOLOv11 in noise resistance and feature representation. RT-DETR significantly reduces missed detections and false alarms in multi-class tasks, achieving superior average precision, recall, and Normalized Root Mean Square Error for parameter extraction. This research offers an efficient approach for intelligent spectrum sensing and non-cooperative reconnaissance in complex environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

RT-DETR-based object detection and parameter extraction for wireless signal spectrograms

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月26日

查看全文
  • Zhibo Shi,
  • Rui Zhu,
  • Lulu Liu,
  • Yaru Li,
  • Hongyan Li,
  • Juan Tian,
  • Le Gao

摘要

Traditional modulation recognition struggles with multi-signal coexistence and low signal-to-noise ratios. This paper proposes an integrated framework for wideband signal detection, recognition, and parameter extraction using object detection technology. By converting signals into spectrograms via Short-Time Fourier Transform, the framework employs RT-DETR as the backbone for feature extraction. Key modules, including Transformer-based Intra-scale Feature Interaction and CNN-based Cross-scale Feature Fusion, are introduced to overcome local convolution limitations, enabling precise end-to-end localization and classification. Furthermore, the system directly extracts physical parameters, such as center frequency and bandwidth, from predicted bounding box geometries. Using a dataset of nine modulation types, experiments across an SNR range of − 30 dB to 30 dB demonstrate that RT-DETR outperforms You Only Look Once (YOLO)v8, YOLOv10, and YOLOv11 in noise resistance and feature representation. RT-DETR significantly reduces missed detections and false alarms in multi-class tasks, achieving superior average precision, recall, and Normalized Root Mean Square Error for parameter extraction. This research offers an efficient approach for intelligent spectrum sensing and non-cooperative reconnaissance in complex environments.