Open Access

ARTICLE

An Edge-Computing-Oriented Small-Object Detection Algorithm for UAV Aerial Images

Chanchan Zhao^1,#, Xiaoyu Gao^1,#, Bao Shi^2,*, Ziyang Zhang¹

1 College of Intelligent Science and Technology, Inner Mongolia University of Technology, Hohhot, China
2 School of Information Engineering, Inner Mongolia University of Technology, Hohhot, China

* Corresponding Author: Bao Shi. Email:
# These authors are co-first authors and contributed equally to this work

Computers, Materials & Continua 2026, 89(1), 66 https://doi.org/10.32604/cmc.2026.085282

Received 08 May 2026; Accepted 02 July 2026; Issue published 13 August 2026

Abstract

Detecting objects in unmanned aerial vehicle (UAV) imagery is challenging because most targets occupy only a small number of pixels and are often distributed in crowded regions with cluttered backgrounds. For edge-side UAV applications, the detector must also remain compact enough for real-time inference on low-power computing platforms. To meet these requirements, this study develops a YOLOv11n-based small-object detector by redesigning feature extraction, cross-scale fusion, and prediction modules. In the backbone, the proposed Dual-Context Large-Small Convolution (DCLSConv) is embedded into the C3k2 structure to form C3k2-DC, allowing the network to capture broader contextual cues while retaining local details useful for small targets. In the neck, Cross-Scale Bi-Attention Fusion (CSBAF) is introduced to regulate the interaction between low-level spatial features and high-level semantic features, thereby reducing ineffective feature mixing during multi-scale fusion. In the detection head, UAV-Task Aligned Dual-Dynamic Detection Head (UAV-TAD²H) is designed to strengthen task-specific feature learning for classification and localization under scale-varying aerial scenes. Experiments on VisDrone2019 show that the proposed model reaches 0.375 mAP@0.5 and 0.230 mAP@0.5:0.95, improving the YOLOv11n baseline by 5.1 and 4.3 percentage points, respectively. The model contains 3.12M parameters and requires 12.3 giga floating-point operations (GFLOPs), while maintaining 33 frames per second (FPS) on Jetson Orin NX under TensorRT FP16 inference. Additional evaluations on PASCAL VOC2012 and BDD100K further indicate that the proposed modules can provide stable performance gains across different detection scenarios.

---

Keywords

---