⚡ electrical engineering

C2D-RT-DETR-Based Wideband Signal Detection and Parameter Extraction

This paper proposes C2D-RT-DETR, a real-time deep learning model that integrates Coordinate Attention, CARAFE, and Dynamic Head mechanisms to achieve robust wideband signal detection and high-precision physical parameter extraction in low signal-to-noise ratio environments, demonstrating superior accuracy and efficiency for intelligent spectrum monitoring.

Original authors: Zhibo Shi, Rui Zhu, Lulu Liu, Yarong Chen, Le Gao, Huanhuan Gheng

Published 2026-09-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhibo Shi, Rui Zhu, Lulu Liu, Yarong Chen, Le Gao, Huanhuan Gheng

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The air around us is thick with invisible waves, a constant hum of radio signals carrying everything from emergency broadcasts to personal messages. In the world of wireless communication, these waves are the lifeblood of modern life, but they are also becoming increasingly crowded. As more devices connect and new technologies emerge, the electromagnetic spectrum is turning into a chaotic jumble where signals overlap, collide, and hide within a sea of static. To manage this, engineers need to listen carefully, identifying exactly what is being transmitted, where it is coming from, and how long it lasts. This task is known as spectrum sensing. However, when the signals are weak or the background noise is loud, traditional listening tools often fail. They struggle to distinguish a faint whisper from the roar of a crowd, leading to missed messages or confused data. The challenge lies in seeing the shape of a signal when it is blurred and broken by interference, a problem that has long hindered the ability to monitor the airwaves effectively.

In a recent study, researchers from Xijing University in China have developed a new way to solve this problem, creating a system that can find and measure these hidden signals even in the most hostile environments. They call their creation C2D-RT-DETR, a sophisticated computer program designed to act like a highly trained eye for radio waves. Instead of relying on old methods that often get confused by static, this new system treats the problem like finding objects in a blurry photograph. It takes the raw, noisy data of radio waves and converts it into a visual map, a two-dimensional image where time runs across the bottom and frequency runs up the side. In this image, a radio signal looks like a bright, colored shape, while the background noise appears as a grainy, chaotic texture. The goal is to draw a precise box around that shape to measure its size, position, and duration, even when the image is heavily distorted.

The researchers found that standard computer vision tools, which are usually good at spotting cars or people in photos, struggle with these radio maps because the signals are often fragmented and the noise is overwhelming. To fix this, they built a custom system that improves how the computer "sees" the image at three critical stages. First, they added a mechanism that helps the computer understand the exact location of a signal by paying close attention to its vertical and horizontal position, much like a person focusing on a specific line of text in a messy paragraph. This step filters out the random static that usually confuses the system. Second, they replaced the standard way the computer enlarges small details with a smarter method that looks at the content of the image before deciding how to fill in the gaps. This prevents the sharp edges of a signal from becoming blurry or smeared when the image is processed. Finally, they added a dynamic control system at the end that separates the tasks of identifying what the signal is from the task of measuring its boundaries, ensuring that the computer does not get confused between the two jobs.

The results of this work are striking, particularly when tested in conditions that mimic real-world chaos. The researchers simulated a wide range of scenarios, from clear conditions to environments where the noise is fifteen times stronger than the signal itself. In these extreme tests, their new system consistently outperformed existing models. When the noise level was moderate, the system correctly identified and measured nearly 93 percent of the signals, a significant improvement over previous methods which often missed a large portion of the targets. Even in the most difficult conditions, where the signal was almost completely buried, the system managed to find the vast majority of the targets, maintaining a high level of accuracy. Beyond just finding the signals, the system could also calculate their physical properties with remarkable precision. It could determine the exact center frequency of a signal and how long it lasted, with errors so small they were barely measurable.

What makes this achievement particularly important is that the system does this without becoming slow or overly complex. Despite the advanced techniques it uses, the program runs quickly enough to be used in real-time applications, processing a new image in less than twenty milliseconds. This speed is crucial for modern monitoring systems that need to react instantly to changing conditions. The researchers also created a vast library of simulated data to train and test their system, covering nine different types of signal patterns and a wide range of noise levels. By validating their approach on this diverse dataset, they demonstrated that their method is not just a theoretical idea but a practical solution that can handle the messy reality of the electromagnetic spectrum. This work offers a reliable path forward for building smarter, more resilient monitoring terminals that can keep the airwaves organized, even when the world around them is noisy and chaotic.

Technical Summary: C2D-RT-DETR-Based Wideband Signal Detection and Parameter Extraction

Problem Statement
In the field of wideband spectrum sensing, accurately detecting weak signals and extracting physical parameters within complex electromagnetic environments remains a significant challenge. Traditional methods, which rely on statistical tools (e.g., higher-order cumulants) or manual time-frequency analysis (e.g., Wigner-Ville Distribution), often suffer from degraded accuracy and poor generalization when facing strong noise interference or dense signal overlaps. While deep learning approaches, particularly the real-time object detection framework RT-DETR, offer a promising alternative by treating spectrum sensing as an image detection problem, they exhibit specific deficiencies in low Signal-to-Noise Ratio (SNR) scenarios. Specifically, standard RT-DETR architectures struggle with:

Feature Smoothing: The use of Nearest Neighbor Upsampling in feature fusion paths tends to smooth out high-frequency details, causing irreversible information loss for weak signals.
Localization Drift: The lack of spatial adaptive mechanisms for time-frequency energy ridges leads to significant deviations in estimating geometric parameters (bandwidth, duration) and physical quantities (center frequency, onset time) under noisy conditions.

Methodology
To address these limitations, the authors propose C2D-RT-DETR, a real-time spectrum sensing model built upon the RT-DETR architecture. The framework integrates three core enhancement mechanisms to achieve high-fidelity reconstruction of corrupted time-frequency features:

Coordinate Attention (CA) Mechanism: Integrated into the shallow feature extraction branches (P3 and P4), this module enhances spatial localization. It employs orthogonal decomposition encoding along the time-delay and frequency axes to capture long-range dependencies while preserving precise coordinate information. This allows the network to anchor signal energy clusters and suppress disordered background noise without losing the spatial data necessary for parameter extraction.
Content-Aware ReAssembly of FEatures (CARAFE): Deployed in the multi-scale feature fusion path, CARAFE replaces the traditional fixed Nearest Neighbor Upsampling. It dynamically generates reconstruction kernels based on local semantic features. This approach avoids the blurring of high-frequency details caused by static interpolation, effectively preserving the edges of weak signals during the upsampling process.
Dynamic Head (DyHead): Integrated at the output stage, DyHead achieves the efficient decoupling of classification and bounding box regression feature flows. Through adaptive regulation in spatial and task dimensions (scale-aware, spatial-aware, and task-aware), it aligns non-stationary, geometrically distorted features, maximizing signal recognizability in the time-frequency domain.

Following detection, the system establishes a mapping mechanism from the geometric boundaries of the detected bounding boxes to the physical domain. This allows for the direct calculation of key physical parameters—center frequency, occupied bandwidth, onset time, and duration—without requiring complex pixel-level segmentation or threshold-based binarization.

Key Contributions
The paper outlines three primary contributions:

Framework Proposal: The introduction of C2D-RT-DETR, an intelligent detection framework that synergizes CARAFE, CoordAtt, and DyHead to overcome the sensing limitations of traditional visual models in low-SNR, multi-signal overlap environments.
Parameter Extraction Model: The refinement and validation of a mathematical mapping model that converts visual bounding box coordinates directly into the physical parameters of wideband signals, offering an engineering solution for automated extraction under harsh channel conditions.
Dataset and Validation: The creation of a wideband time-frequency dataset containing 9 complex modulation types across an SNR range of -15 dB to 5 dB. Extensive comparative and ablation experiments were conducted to validate the model's efficacy.

Experimental Results
The proposed model was evaluated against the baseline RT-DETR-L and mainstream YOLO series (v8, v10, v11) models. Key findings include:

Detection Performance: In a 0 dB noise environment, C2D-RT-DETR achieved a mean Average Precision ( $mAP_{50}$ ) of 0.687, representing a 3.1 percentage point improvement over the baseline RT-DETR-L (0.656). The model demonstrated robust anti-noise capabilities, maintaining a detection accuracy of 0.612 even at -15 dB SNR.
Target Recall: The model achieved a target matching coverage rate of up to 92.8% in 0 dB environments, significantly outperforming YOLO variants which suffered from high miss rates due to signal fragmentation.
Parameter Estimation Accuracy: The Normalized Root Mean Square Errors (NRMSE) for critical parameters were exceptionally low: 0.018% for center frequency and 0.157% for onset time. Bandwidth and duration estimation errors remained under 3% in moderate noise conditions.
Efficiency: The model maintains a lightweight footprint with 32.76 M parameters and 105.0 GFLOPs. The single-frame inference latency is 19.46 ms (approx. 51.4 FPS), satisfying the real-time requirements for spectrum monitoring while balancing computational overhead.

Significance
The paper claims that C2D-RT-DETR provides a reliable technical solution for the intelligent deployment of non-cooperative wideband spectrum monitoring terminals. By establishing a robust link between visual geometric boundaries and physical signal parameters, the research offers a practical engineering approach to achieving high-precision, real-time spectrum sensing in complex electromagnetic environments where traditional methods and standard deep learning models fail. The work demonstrates that a synergistic combination of coordinate sensing, content-aware reassembly, and dynamic task alignment can effectively bridge the gap between visual object detection and physical signal analysis under extreme noise conditions.

Technical Summary: C2D-RT-DETR-Based Wideband Signal Detection and Parameter Extraction

More like this