Scholarlink

<p>To address the challenges in Unmanned Aerial Vehicle (UAV) aerial image object detection, such as high computational complexity, deployment difficulties, small-object sizes, and severe background interference, this study proposes a Lightweight and Efficient UAV Aerial Image Small Object Detection model (LEUS-YOLO) based on YOLO11n. First, a Re-parameterized Feature Extraction module (Rep2C3k2) is proposed to initially extract multi-scale features of objects in aerial images by leveraging the advantages of structural re-parameterization and cross-stage partial network structures. Second, a Multi-Scale Cross-Stage Feature Fusion module (MSCS) is designed to significantly enhance the model’s feature representation capability through shallow feature fusion branches and multi-level feature processing with a decoupled attention mechanism. Third, the Lightweight Shared Convolution Detection Head (LS-Head) effectively improves detection efficiency and reduces model complexity through parameter sharing and hybrid normalization strategies. Finally, a Spatial-Channel Downsampling module (SCDown) is introduced between the backbone and neck networks to optimize feature downsampling and further enhance lightweight performance. To verify the effectiveness of the proposed method, experiments were conducted on the VisDrone2019 and DOTAv1.5 datasets. On the VisDrone2019 dataset, compared with YOLO11n, LEUS-YOLO achieved improvements of 2.8% and 1.9% in mAP<sub>50</sub> and mAP<sub>50:95</sub>, respectively, while reducing the number of parameters by 59.7%, lowering the computational cost by 9.5%, compressing the model size by 51.1%, and increasing FPS by 10.9%. This model significantly enhances the feasibility of deployment on edge devices and improves real-time performance, demonstrating broader application prospects in UAV aerial image object detection.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LEUS-YOLO: a lightweight and efficient UAV-based object detection model for small targets in aerial images

  • 期刊论文

  • 发表日期：

    2026年4月2日

查看全文
  • Dingxuan Sheng,
  • Yecai Guo

摘要

To address the challenges in Unmanned Aerial Vehicle (UAV) aerial image object detection, such as high computational complexity, deployment difficulties, small-object sizes, and severe background interference, this study proposes a Lightweight and Efficient UAV Aerial Image Small Object Detection model (LEUS-YOLO) based on YOLO11n. First, a Re-parameterized Feature Extraction module (Rep2C3k2) is proposed to initially extract multi-scale features of objects in aerial images by leveraging the advantages of structural re-parameterization and cross-stage partial network structures. Second, a Multi-Scale Cross-Stage Feature Fusion module (MSCS) is designed to significantly enhance the model’s feature representation capability through shallow feature fusion branches and multi-level feature processing with a decoupled attention mechanism. Third, the Lightweight Shared Convolution Detection Head (LS-Head) effectively improves detection efficiency and reduces model complexity through parameter sharing and hybrid normalization strategies. Finally, a Spatial-Channel Downsampling module (SCDown) is introduced between the backbone and neck networks to optimize feature downsampling and further enhance lightweight performance. To verify the effectiveness of the proposed method, experiments were conducted on the VisDrone2019 and DOTAv1.5 datasets. On the VisDrone2019 dataset, compared with YOLO11n, LEUS-YOLO achieved improvements of 2.8% and 1.9% in mAP50 and mAP50:95, respectively, while reducing the number of parameters by 59.7%, lowering the computational cost by 9.5%, compressing the model size by 51.1%, and increasing FPS by 10.9%. This model significantly enhances the feasibility of deployment on edge devices and improves real-time performance, demonstrating broader application prospects in UAV aerial image object detection.