Frequency-aware Mamba and dynamic optimization DETR for real-time tiny-object detection in UAV aerial imagery

Abstract

实时微型物体探测是低空无人飞行器(UAV)感知的关键任务。然而,由于成像高度、视角和复杂背景的变化,空中目标通常体积小、分布密集且视觉模糊,从而对检测精度和推理效率构成了巨大挑战。尽管RT-DETR框架提供了端到端的检测模式,但它对微小物体的反应较弱,特征融合过程中细微细节的退化,以及无人机场景中不稳定的边界箱回归。为应对这些限制,我们的FAMO-DETR被引入为一种频率感知的端到端探测器,专为通过无人机探测微小物体而设计。具体而言,通过离散的小波变换,将频率感知的Mamba编码器(FAME)设计用于将低频语义信息与高频边缘细节分离,从而通过Mamba序列建模,增强深度特征中的全局情境表示。随后开发出一种高频CARAFE模块(HF-CARAFE),将高频边缘引导和动态内核重新组装为特征融合,从而缓解传统升级操作引起的语义错位和边界模糊。最后,开发了NWD-CIoU损耗(NC-Loss),将归一化的Wasserstein距离的平滑距离约束与CIOU的几何重叠约束相结合,为微小物体定位提供了更稳定的优化信号。对VisDrone2019-DET和AI-TOD数据集的全面实验表明,FAMO-DETR实现了 AP_s 在VisDrone2019-DET上,数值为22.1%,在AI-TOD上为AP值为23.5%,使基准配置分别提高了1.7%和2.3%,同时保持了134帧率的实时推理速度。

Target-Guided Query Allocation and Boundary-Aware Multi-Scale Feature Fusion for Object Detection in UAV Imagery
Sensors (Basel, Switzerland)26 Sep 2026Peer-reviewedCC BYFull textEurope PMC
MSAL-YOLO: a YOLOv8-based detector for small and densely distributed object detection in UAV aerial imagery
Fei Ding, Xiufu Du, Haining Zhang +2
Scientific Reports7 Oct 2026Peer-reviewedCC BY-NC-ND 4.0Full textCrossref
MF-YOLO: High-Resolution Weighted Feature Fusion with Global-Local Aggregation for Small-Object Detection in UAV Imagery
Yufei Wang, Jiming Yang, Fang Liu +1
Sensors (Basel, Switzerland)25 Sep 2026Peer-reviewedCC BYFull textEurope PMC
Comparative Evaluation of Deep Object Detection Architectures for Automatic Target Recognition in Real-World AUV-Acquired Side-Scan Sonar Imagery
Daniel Powarzyński, Norbert Sigiel, Aleksander Cirocki +1
Sensors (Basel, Switzerland)1 Oct 2026Peer-reviewedCC BYFull textEurope PMC
Differentiable Resolution Allocation for Tiny Object Detection in UAV Imagery
Ergashevich Halimjon Khujamatov, Khasanov Doston, Yusupov Sarvarbek Sodikovich +5
Sensors (Basel, Switzerland)28 Sep 2026Peer-reviewedCC BYFull textEurope PMC

Cite

BibTeX RIS

BibTeX

@article{zhang2026frequencyaware,
  title = {{Frequency-aware Mamba and dynamic optimization DETR for real-time tiny-object detection in UAV aerial imagery}},
  author = {Sen Zhang and Chonghang Wang and Chenyang Li and Zhuangzhuang Zhou and Dengxi Chen and Shukui Bo and Rong Nie},
  journal = {Scientific Reports},
  year = {2026},
  month = oct,
  doi = {10.1038/s41598-026-73297-x},
  url = {https://www.nature.com/articles/s41598-026-73297-x},
  note = {CC BY 4.0},
}

RIS

TY  - JOUR
TI  - Frequency-aware Mamba and dynamic optimization DETR for real-time tiny-object detection in UAV aerial imagery
AU  - Sen Zhang
AU  - Chonghang Wang
AU  - Chenyang Li
AU  - Zhuangzhuang Zhou
AU  - Dengxi Chen
AU  - Shukui Bo
AU  - Rong Nie
PY  - 2026
DA  - 2026/10/06
JO  - Scientific Reports
DO  - 10.1038/s41598-026-73297-x
UR  - https://www.nature.com/articles/s41598-026-73297-x
AB  - 实时微型物体探测是低空无人飞行器(UAV)感知的关键任务。然而,由于成像高度、视角和复杂背景的变化,空中目标通常体积小、分布密集且视觉模糊,从而对检测精度和推理效率构成了巨大挑战。尽管RT-DETR框架提供了端到端的检测模式,但它对微小物体的反应较弱,特征融合过程中细微细节的退化,以及无人机场景中不稳定的边界箱回归。为应对这些限制,我们的FAMO-DETR被引入为一种频率感知的端到端探测器,专为通过无人机探测微小物体而设计。具体而言,通过离散的小波变换,将频率感知的Mamba编码器(FAME)设计用于将低频语义信息与高频边缘细节分离,从而通过Mamba序列建模,增强深度特征中的全局情境表示。随后开发出一种高频CARAFE模块(HF-CARAFE),将高频边缘引导和动态内核重新组装为特征融合,从而缓解传统升级操作引起的语义错位和边界模糊。最后,开发了NWD-CIoU损耗(NC-Loss),将归一化的Wasserstein距离的平滑距离约束与CIOU的几何重叠约束相结合,为微小物体定位提供了更稳定的优化信号。对VisDrone2019-DET和AI-TOD数据集的全面实验表明,FAMO-DETR实现了 AP_s 在VisDrone2019-DET上,数值为22.1%,在AI-TOD上为AP值为23.5%,使基准配置分别提高了1.7%和2.3%,同时保持了134帧率的实时推理速度。
SN  - 2045-2322
ER  -

Abstract

Related

Cite