Scholarlink

<p>Intelligent transportation has emerged as a prominent research focus, with object detection playing a critical role in its development. While existing models achieve robust results for medium and large objects, small objects in vehicle scenarios remain challenging, significantly increasing detection difficulty. Addressing this challenge, this paper proposes LS-YOLO, a lightweight small-object detection algorithm based on YOLOv11. First, to improve small-object detection without compromising performance on medium/large objects, a scale-adaptive dynamic IoU loss (SADIoU) is introduced. By dynamically adjusting the scaling factor, this approach ensures precise regression for small objects while maintaining accuracy for larger targets. Second, a self-attention mechanism module (SHSA) is incorporated to enhance feature extraction and fusion capabilities. Additionally, the neck structure of YOLO is improved by the proposal of LSFE-PAFPN (Lightweight Small-object Feature Extraction Feature Pyramid Network), strengthening shallow feature utilization for micro-object detection. LS-YOLO was trained and tested on the VisDrone2019 dataset, which contains numerous small objectscommon in vehicle scenarios and features a high proportion of small targets. Compared to YOLOv11, LS-YOLO achieved a 5.4% improvement in mAP0.5 on VisDrone2019 while reducing parameter count by 20.9% and requiring lower hardware resources. These results demonstrate LS-YOLO’s effectiveness in balancing detection accuracy and computational efficiency for small-object scenarios in intelligent transportation systems.</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

LS-YOLO: a lightweight small-object detection framework with region scaling loss and self-attention for intelligent transportation systems

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Yingxiao Yu,
  • Minghao Huang,
  • Kun Wang,
  • Xiangyu Tang,
  • Junjie Bao,
  • Yukun Fan

摘要

Intelligent transportation has emerged as a prominent research focus, with object detection playing a critical role in its development. While existing models achieve robust results for medium and large objects, small objects in vehicle scenarios remain challenging, significantly increasing detection difficulty. Addressing this challenge, this paper proposes LS-YOLO, a lightweight small-object detection algorithm based on YOLOv11. First, to improve small-object detection without compromising performance on medium/large objects, a scale-adaptive dynamic IoU loss (SADIoU) is introduced. By dynamically adjusting the scaling factor, this approach ensures precise regression for small objects while maintaining accuracy for larger targets. Second, a self-attention mechanism module (SHSA) is incorporated to enhance feature extraction and fusion capabilities. Additionally, the neck structure of YOLO is improved by the proposal of LSFE-PAFPN (Lightweight Small-object Feature Extraction Feature Pyramid Network), strengthening shallow feature utilization for micro-object detection. LS-YOLO was trained and tested on the VisDrone2019 dataset, which contains numerous small objectscommon in vehicle scenarios and features a high proportion of small targets. Compared to YOLOv11, LS-YOLO achieved a 5.4% improvement in mAP0.5 on VisDrone2019 while reducing parameter count by 20.9% and requiring lower hardware resources. These results demonstrate LS-YOLO’s effectiveness in balancing detection accuracy and computational efficiency for small-object scenarios in intelligent transportation systems.