FUR-DETR: A Lightweight Detection Model for Fixed-Wing UAV Recovery

Yu Yao, Jun Wu, Yongchang Hao, Zhen Huang, Zixuan Yin, Jiajing Xu, Honglin Chen, Jin‐Hong PiDrones 2025 · DOI

TL;DR

This paper introduces FUR-DETR, a lightweight Transformer-based object detection model designed for fixed-wing UAV recovery. By proposing a novel Parallel Backbone, an Adaptive Multiscale Feature Pyramid Network (AMFPN), and a Wavelet Transform-based Frequency-Optimized Reverse Fusion Mechanism (WT-FORM), FUR-DETR significantly reduces model size and computational load while maintaining detection accuracy comparable to existing state-of-the-art methods, even in challenging conditions.

What Problem Does This Paper Address?

Traditional UAV recovery systems often struggle in communication-constrained or GPS-denied environments, lacking real-time visual perception for decision-making and parameter adjustments. This increases recovery risks and limits autonomy. While visual inspection offers a solution, existing deep learning models like RT-DETR face limitations: single-path feature extraction, simplified fusion, and loss of high-frequency information, hindering their efficiency and accuracy, especially for small targets in complex scenarios (e.g., varying light, occlusion). The paper aims to bridge this gap by developing a lightweight, efficient, and robust object detection model specifically for UAV recovery.

Methods

The core of FUR-DETR lies in three novel components:

  • Parallel Backbone: This dual-branch architecture extracts local and global semantic features in parallel. It uses a shared initial module and then branches out, with one path focusing on fine-grained details (inspired by CSPNet) and the other on semantic enhancement. A progressive fusion mechanism adaptively integrates these multiscale features.
  • Adaptive Multiscale Feature Pyramid Network (AMFPN): This component enhances feature fusion and upsampling. It employs a hierarchical information transmission and two-way fusion strategy to combine high-level semantics with low-level spatial details. An adaptive fusion mechanism dynamically weights features from different scales, improving representation and reducing complexity.
  • Wavelet Transform-based Frequency-Optimized Reverse Fusion Mechanism (WT-FORM): This module addresses information loss during downsampling. It uses wavelet transform to decompose shallow features into low-frequency (global) and high-frequency (local) components. These components are then adaptively reweighted and fused, preserving high-frequency details crucial for small target detection and expanding the receptive field.

Datasets

The authors evaluated their model on three public UAV datasets:

  • Det-Fly: An air-to-air UAV detection dataset simulating various conditions like lighting, blur, and occlusion.
  • DUT-Anti-UAV: A dataset focusing on UAV detection and tracking, with multiple resolutions to test multiscale adaptability.
  • TIB: A ground-to-air dataset primarily consisting of small-target images.

Key Insights and Ideas

  • Dual-branch architecture is key: The Parallel Backbone demonstrates that splitting feature extraction into parallel paths (fine-grained vs. semantic) can capture richer information and improve efficiency compared to single-path networks.
  • Frequency domain matters for small targets: WT-FORM highlights the importance of leveraging frequency domain information (beyond just spatial domain optimization) to retain high-frequency details essential for detecting small and occluded targets.
  • Synergy of modules: The paper strongly emphasizes that the combined effect of Parallel Backbone, AMFPN, and WT-FORM is greater than the sum of their parts, leading to significant improvements in both efficiency and accuracy.
  • Lightweight doesn't mean less accurate: FUR-DETR proves that substantial reductions in model parameters and computational load are achievable without sacrificing, and sometimes even improving, detection performance.

Key Results

  • Parameter and Computation Reduction: FUR-DETR reduces parameter size by 43.2% and computational load (GFLOPS) by 58% compared to the baseline RT-DETR.
  • Comparable Accuracy: The model maintains detection accuracy comparable to the original RT-DETR across three datasets.
  • Improved Performance in Challenging Conditions: FUR-DETR shows superior performance in low light, occlusion, and small-target scenarios compared to RT-DETR and other baseline models.
  • Specific Dataset Performance: On the Det-Fly dataset, FUR-DETR achieved an AP50-95 of 0.662 with only 11.3M parameters and 23.9 GFLOPS, outperforming many other lightweight models. On DUT-Anti-UAV, it achieved an AP50-95 of 0.698, and on TIB, an AP50-95 of 0.375.

Limitations and Future Work

  • Out-of-Distribution Samples: The authors acknowledge that deep learning models can degrade with out-of-distribution samples. While FUR-DETR performs well on public datasets, its performance in extreme lighting or severe weather conditions needs further investigation.
  • Extreme Conditions: Future work will focus on improving performance under extreme lighting and severe weather.
  • Adaptability: Addressing challenges like occlusion and motion blur will be a key area for future research.
  • Small Sample & Meta-Learning: The authors plan to explore small sample and meta-learning methods to enhance the model's adaptability to new environments with minimal training data.
  • Hardware Optimization: Developing hardware-oriented optimization strategies is also planned to improve performance on deployment platforms.

Summary updated

Abstract

Due to traditional recovery systems lacking visual perception, it is difficult to monitor UAVs’ real-time status in communication-constrained or GPS-denied environments. This leads to insufficient ability in decision-making and parameter adjustment and increase uncertainty and risk of recovery. Visual inspection technology can make up for the limitations of GPS and communication and improve the autonomy and adaptability of the system. However, the existing RT-DETR algorithm is limited by single-path feature extraction, a simplified fusion mechanism, and high-frequency information loss, which makes it difficult to balance detection accuracy and computational efficiency. Therefore, this paper proposes a lightweight visual detection model based on transformer architecture to further optimize computational efficiency. Firstly, aiming at the performance bottleneck of existing models, the Parallel Backbone is proposed, which captures local features and global semantic information by sharing the initial feature extraction module and the double-branch structure, respectively, and uses the progressive fusion mechanism to realize the adaptive integration of multiscale features so as to balance the accuracy and lightness of target detection. Secondly, an adaptive multiscale feature pyramid network (AMFPN) is designed, which effectively integrates different scales of information through multi-level feature fusion and information transmission mechanism, alleviates the problem of information loss in small-target detection, and improves the detection accuracy in complex backgrounds. Finally, a wavelet frequency–domain-optimized reverse feature fusion mechanism (WT-FORM) is proposed. By using the wavelet transform to decompose the shallow features into multi-frequency bands and combining the weighted calculation and feature compensation strategy, the computational complexity is reduced, and the representation ability of the global context is further enhanced. The experimental results show that the improved model reduces the parameter size and computational load by 43.2% and 58% while maintaining detection accuracy comparable to the original RT-DETR in three datasets. Even in complex environments with low light, occlusion, or small targets, it can provide more accurate detection results.
BibTeX
@inproceedings{yao2025fur,
  author = {Yu Yao and Jun Wu and Yongchang Hao and Zhen Huang and Zixuan Yin and Jiajing Xu and Honglin Chen and Jin‐Hong Pi},
  title = {{FUR-DETR: A Lightweight Detection Model for Fixed-Wing UAV Recovery}},
  booktitle = {Drones},
  year = {2025},
  doi = {10.3390/drones9050365},
}
Figure