DSAB-FPN: Role-Separated High-Resolution Neck Fusion for UAV Small-Object Detection
IEEE Access · 2026 · 14 · 168985-168997
DOI: 10.1109/access.2026.3748090 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Young-Sun Ahn und 2 weiteren Autor:innen. Die Arbeit nennt 29 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Basic Science Research Program through the National Research Foundation of Korea, Ministry of Education, National Research Foundation of Korea (NRF) grants.
Abstract (Originalsprache)
Small-object detection in unmanned aerial vehicle (UAV) imagery remains challenging because targets occupy few pixels and often appear in dense, occluded, and cluttered scenes. Adding a stride-4 P2 detection level preserves spatial detail, but a uniform fusion rule along the top–down path does not account for differences in source relationships. This paper proposes the Detail–Semantic Aggregation Bridge Feature Pyramid Network (DSAB-FPN), a compact high-resolution neck for YOLOv8n-P2 that assigns fusion mechanisms according to node role. Detail–Semantic Fusion (DSFuse) uses fixed priors selected for each node when fusing adjacent sources, whereas the Detail–Semantic Bridge (DSB) adaptively balances the sources at the intermediate bridge. Controlled experiments on VisDrone2019-DET show consistent improvements in both overall and small-object detection accuracy. DSAB-FPN achieves an average precision (AP) of 0.2392, improving the direct YOLOv8n-P2 baseline by 0.0199 and standard YOLOv8n by 0.0446. Component ablations support complementary roles for the two fusion mechanisms, and independent evaluation on Berkeley DeepDrive 100K (BDD100K) shows the same improvement trend under a different scene distribution. The final model contains 3.0864 million parameters, requires 14.23 billion floating-point operations, and reaches 121.52 frames/s during TensorRT measurements of network execution on a Jetson Orin Nano Super in its maximum performance mode. These results indicate that role-separated fusion improves small-object detection while quantifying the balance between accuracy and throughput on an edge device.
Förderung
- Basic Science Research Program through the National Research Foundation of Korea
- Ministry of Education
- National Research Foundation of Korea (NRF) grants
- Korean Government, the Ministry of Science and ICT
- Daegu Gyeongbuk Institute of Science and Technology (DGIST) Research and Development Program of the Ministry of Science and ICT
Zitieren
APA
Ahn, Y., Lee, C., & Yang, J. (2026). DSAB-FPN: Role-Separated High-Resolution Neck Fusion for UAV Small-Object Detection. IEEE Access, 14, 168985-168997. https://doi.org/10.1109/access.2026.3748090
Vancouver
Ahn Y, Lee C, Yang J. DSAB-FPN: Role-Separated High-Resolution Neck Fusion for UAV Small-Object Detection. IEEE Access. 2026;14:168985-168997. doi:10.1109/access.2026.3748090
BibTeX
@article{Ahn2026,
author = {Ahn, Young-Sun and Lee, Chung-Hee and Yang, Jung-Min},
title = {DSAB-FPN: Role-Separated High-Resolution Neck Fusion for UAV Small-Object Detection},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {168985--168997},
doi = {10.1109/access.2026.3748090}
}
EUVIMED