Scholarlink

<p>Object detection in UAV images is a challenging task in computer vision. Objects in UAV images have difficulties such as low resolution, large-scale variations, and complex backgrounds. In this study, we propose a novel real-time detection model based on YOLO11 to solve the above problems. First, we created a better feature extraction module to better extract the small-object features. Second, we designed a feature pyramid that improves the model detection by using local and global information of objects. In addition, we design a task interaction detection header that improves model localization and classification through task alignment. We also developed Inner-Wise-MPDIoUv2 to address the limitations of CIoU in detecting small objects. Finally, we use model pruning to reduce the model size. SOO-YOLO achieves a 6.1% improvement over YOLO11n on the VisDrone2019-DET dataset’s <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\text {AP}_\text {.50}\)</EquationSource> </InlineEquation> score despite its compact size of just 0.78M. It also demonstrates excellent performance on the DOTA dataset.</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

SOO-YOLO: an efficient small object detection model for UAV images

  • 期刊论文

  • 发表日期：

    2026年1月6日

查看全文
  • Renjie Chen,
  • Hua Sun,
  • Haiyang Fan,
  • Pingxiang Wu,
  • Zhenqi Zhang,
  • Yang Chen

摘要

Object detection in UAV images is a challenging task in computer vision. Objects in UAV images have difficulties such as low resolution, large-scale variations, and complex backgrounds. In this study, we propose a novel real-time detection model based on YOLO11 to solve the above problems. First, we created a better feature extraction module to better extract the small-object features. Second, we designed a feature pyramid that improves the model detection by using local and global information of objects. In addition, we design a task interaction detection header that improves model localization and classification through task alignment. We also developed Inner-Wise-MPDIoUv2 to address the limitations of CIoU in detecting small objects. Finally, we use model pruning to reduce the model size. SOO-YOLO achieves a 6.1% improvement over YOLO11n on the VisDrone2019-DET dataset’s \(\text {AP}_\text {.50}\) score despite its compact size of just 0.78M. It also demonstrates excellent performance on the DOTA dataset.