SC-D-FINE: Early Multi-Scale Focusing and Context-Aware Bridging for Lightweight Object Detection in Aerial Imagery
IEEE Access · 2026 · 14 · 120435-120450
DOI: 10.1109/access.2026.3713646 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Shuhang Tang und 1 weiteren Autor:innen. Die Arbeit nennt 40 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Science and Technology Research Program of Chongqing Municipal Education Commission.
Abstract (Originalsprache)
Lightweight object detection in aerial imagery is important for resource-constrained unmanned aerial vehicles (UAVs) and edge platforms used in traffic monitoring, aerial inspection, and surveillance. Under tight computational budgets below 15 giga floating-point operations (GFLOPs), aggressive simplification of the backbone–neck pipeline often weakens contextual modeling and cross-scale interaction, which is particularly detrimental in aerial scenes dominated by small and crowded objects. We present SC-D-FINE, a lightweight D-FINE-based Detection Transformer (DETR)-style detector for aerial imagery. The design centers on an SGMA–CaMPB encoder for early multi-scale enhancement, with a re-parameterizable DRC tail providing low-cost high-level context. In the encoder, SGMA injects adjacent-scale cues, whereas CaMPB performs compact feature alignment and bidirectional context fusion to reduce scale-induced noise. On VisDrone2019, SC-D-FINE-N reaches 25.7% Average Precision (AP) and 43.0% AP50 (AP at an Intersection over Union threshold of 0.50), compared with 20.8% AP and 36.8% AP50 for D-FINE-N. In the 14-GFLOPs regime, it remains 1.8 AP above the wider D-FINE-N-14G variant while using 30.1% fewer parameters. On UAVVaste and Vehicle Detection in Aerial Imagery (VEDAI), the same design yields AP gains of 6.4 and 8.4 points over D-FINE-N. Overall, these results suggest that the gains, driven mainly by the encoder changes and supported by the lightweight DRC tail, are not confined to VisDrone2019.
Förderung
- Science and Technology Research Program of Chongqing Municipal Education Commission
Zitieren
APA
Tang, S., & Zeng, Z. (2026). SC-D-FINE: Early Multi-Scale Focusing and Context-Aware Bridging for Lightweight Object Detection in Aerial Imagery. IEEE Access, 14, 120435-120450. https://doi.org/10.1109/access.2026.3713646
Vancouver
Tang S, Zeng Z. SC-D-FINE: Early Multi-Scale Focusing and Context-Aware Bridging for Lightweight Object Detection in Aerial Imagery. IEEE Access. 2026;14:120435-120450. doi:10.1109/access.2026.3713646
BibTeX
@article{Tang2026,
author = {Tang, Shuhang and Zeng, Zhuo},
title = {SC-D-FINE: Early Multi-Scale Focusing and Context-Aware Bridging for Lightweight Object Detection in Aerial Imagery},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {120435--120450},
doi = {10.1109/access.2026.3713646}
}
EUVIMED