Scholarlink

<p>The swift advancement of drone technology has highlighted the essential role of precise object detection in drone images across a variety of use cases, including remote sensing, traffic monitoring, and military reconnaissance. However, existing detection methods often struggle with challenges like small target detection, complex backgrounds, and multi-scale feature fusion. In order to overcome these challenges, we introduce UCS-YOLO, a refined model built upon YOLOv8. The UCS-YOLO framework incorporates several key advancements: (1) an improved ConvFormer module, adapted from the MetaFormer architecture, to augment the C2f module for detailed feature extraction; (2) a new neck architecture integrating CSPCARS with SPDConv modules for efficient fusion of features at multiple scales. The CSPCARS component utilizes channel and spatial attention mechanisms to emphasize key features, while the SPDConv module enhances feature fusion and context awareness across varying scales and environments. Additionally, we propose the ShapeAware IoU loss function, which refines traditional IoU metrics by integrating overlap ratio, center deviation, and shape adaptation, ensuring more accurate small target detection. On the VisDrone2019 dataset, UCS-YOLOv8 achieves a 3.08% improvement in detection accuracy compared to the baseline model. This demonstrates the model’s superior capability in detecting small targets and its effectiveness in handling complex scenes.</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

UCS-YOLO: a novel approach for multi-scale small target detection in UAVs

  • 期刊论文

  • 发表日期：

    2025年6月30日

查看全文
  • Guang-Ling Sun,
  • Fen-Qi Zhang,
  • Yu-Min Zhu,
  • Fei Miao

摘要

The swift advancement of drone technology has highlighted the essential role of precise object detection in drone images across a variety of use cases, including remote sensing, traffic monitoring, and military reconnaissance. However, existing detection methods often struggle with challenges like small target detection, complex backgrounds, and multi-scale feature fusion. In order to overcome these challenges, we introduce UCS-YOLO, a refined model built upon YOLOv8. The UCS-YOLO framework incorporates several key advancements: (1) an improved ConvFormer module, adapted from the MetaFormer architecture, to augment the C2f module for detailed feature extraction; (2) a new neck architecture integrating CSPCARS with SPDConv modules for efficient fusion of features at multiple scales. The CSPCARS component utilizes channel and spatial attention mechanisms to emphasize key features, while the SPDConv module enhances feature fusion and context awareness across varying scales and environments. Additionally, we propose the ShapeAware IoU loss function, which refines traditional IoU metrics by integrating overlap ratio, center deviation, and shape adaptation, ensuring more accurate small target detection. On the VisDrone2019 dataset, UCS-YOLOv8 achieves a 3.08% improvement in detection accuracy compared to the baseline model. This demonstrates the model’s superior capability in detecting small targets and its effectiveness in handling complex scenes.