Related Experiment Videos
SPA-DETR: An Enhanced RT-DETR with Spatial-Preserving Attention and Adaptive Loss for UAV Spectrogram Signal
Conghao Fu1, Lu Xu1, Yijia Zhang1
1School of Information Science and Engineering, Zhejiang Sci-Tech University, Hangzhou 310018, China.
Sensors (Basel, Switzerland)
|August 13, 2026
Summary
We developed SPA-DETR, a novel deep learning model for detecting unauthorized drones using radio frequency spectrograms. This efficient system enhances low-altitude security by accurately identifying weak signals amidst background noise.
Area of Science:
- Computer Vision
- Signal Processing
- Artificial Intelligence
Background:
- Unauthorized drones pose significant low-altitude security risks.
- Detecting transient, frequency-hopping drone signals in RF spectrograms is challenging for standard object detectors due to signal loss and background noise.
- Existing methods struggle with the microscopic spatial footprints of weak signals.
Purpose of the Study:
- To propose SPA-DETR, a custom deep learning architecture for enhanced detection of unmanned aerial vehicles (UAVs) in radio frequency (RF) spectrograms.
- To improve the accuracy and efficiency of real-time drone detection for low-altitude security applications.
- To address limitations in preserving spatial details of weak signals and mitigating background noise.
Main Methods:
- Introduced the Spatial-Preserving Attention (SPA) block, integrating Space-to-Depth Convolution (SPDConv) and Parallel Patch-Aware Attention (PPA), to retain spatial details and filter noise.
- Developed an Adaptive Threshold Focal Loss (ATFL) to manage foreground-background imbalance during training, focusing the network on difficult signal patches.
- Utilized the RT-DETR framework as the base architecture for SPA-DETR.
Main Results:
- SPA-DETR achieved a mean Average Precision (mAP50:95) of 86.8% and an APS of 85.6% on the RFUAV dataset.
- Demonstrated significant improvements over the baseline RT-DETR-R18 by 4.9% (mAP50:95) and 5.2% (APS).
- Achieved high performance with 235.2 FPS and only 23.74 million parameters, outperforming models like YOLOv10m, YOLOv8m, and RT-DETR-R50.
Conclusions:
- SPA-DETR effectively detects unauthorized UAVs in RF spectrograms by preserving spatial details and reducing background noise.
- The proposed architecture offers a computationally efficient and practical solution for real-time low-altitude security.
- SPA-DETR represents a significant advancement in drone detection technology for security applications.
