Multiscale Attention-enchanced Yolo: a Hierarchical Feature Fusion Framework for Uav Object Detection in Low-illumination Environments
Year: 2025
Type: conference-paper
Abstract: Low-light object detection is a critical computer vision task in UAV applications, as insufficient illumination significantly degrades image quality and subsequently weakens object detection and localisation performance. To address this challenge, a cascaded enhancement–detection framework is proposed. Firstly, a dual-domain guided real-time low-light enhancement network, DDNet, is employed to restore colour and gradient information of dark images in the dataset, thereby improving detail visibility. An improved YOLOv11 detector is then applied to the enhanced dataset for object detection. Specifically, a Multiscale Kernel Decomposition Attention module is embedded in the C3k2 layer of YOLOv11 to extract more effective multiscale features. In addition, Spatial–Channel Synergistic Attention modules are inserted at key fusion points between the Neck and Head networks to alleviate semantic bias in multiscale features and guide the network to more accurately focus on target regions. Experimental results demonstrate that the proposed method achieves a performance of 21.5% on the low-light subset of the VisDrone2019 dataset, representing a 3.4 percentage point improvement over the YOLOv11 baseline, thereby providing a deployable solution for UAV vision tasks conducted at night or under extreme lighting conditions.
Authors: Jiaqi Wang, Liwen Liu, Gui Fu, Xiaoqiang Zhang
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7116417536
