Recieved:

20/07/2026

Accepted:

15/09/2026

Page: 

–

doi:

http://dx.doi.org/10.17515/resm2026-1923vk0720rs

Views:

21

High-Frequency detail injection and normalized Wasserstein distance for small object detection in UAV remote sensing using YOLOv8

Qifeng Han 1, Mingmin Liu 1, Xiaomin Han1, Yuying Ma 1

1School of Electronic Engineering, Shandong University of Engineering and Vocational Technology, Jinan, China

Abstract

To address feature attenuation, scale imbalance, and background interference in UAV remote sensing images, an enhanced YOLOv8 framework is proposed. A high-frequency detail enhancement (HF) module preserves shallow texture information, adaptive feature fusion (AFF) improves cross-scale representation, and global context attention (GCA) strengthens semantic consistency. Normalized Wasserstein Distance (NWD) is introduced to improve small-object localization stability. On the VisDrone-DET2021 validation set, mAP@0.5 increases from 76.2% to 81.6%, while mAP-S increases from 48.7% to 54.5%. The model contains 15.3 M parameters and requires 38.9 GFLOPs, representing increases of 36.6% and approximately 36.0%, respectively, relative to YOLOv8s. It maintains an inference speed of 88.2 FPS at an input resolution of 640×640.

Keywords

UAV remote sensing; Small object detection; YOLOv8; Normalized Wasserstein distance; High-Frequency detail enhancement

Cite this article as: 

Share This Article
LinkedIn
X
Facebook
journal cover
News & Upcoming Events