Scholarlink

<p>With the continuous development of object detection technology and drone aerial photography technology, object detection technology is increasingly applied in the process of detecting targets captured by drones. However, in the perspective of drones, there will be many small targets, occlusions, and a large number of complex environments. Therefore, this paper proposes the CSF-YOLOv8 small target detection algorithm. The algorithm first proposed CSPPC to replace the C2f in the backbone of YOLOv8, in order to enhance the model’s adaptability to scale changes and its ability to detect small-scale objects. SENetV2 has been added to the output of the small object detection layer in the neck network. This design aims to improve the precision of feature expression and the ability to integrate global information through a multi branch structure, enhancing the algorithm’s detection ability in complex situations and facing targets of different scales. The use of FRMHead detection head has replaced the original detection head, improving the accuracy of object detection in the model. Modify the original loss function CIoU to InnerWIoU, which introduces auxiliary bounding boxes of different scales to calculate the loss. This loss function is suitable for detecting targets with uneven distribution or large size differences. Experimental data shows that the improved algorithm has an 8.4% increase in mAP50 compared to the original algorithm on the VisDrone2019 dataset, mAP50%: 95% increased by 5.9%.</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

Small target detection algorithm based on CSF-YOLOv8

  • 期刊论文

  • 发表日期：

    2025年7月30日

查看全文
  • Xiaoxia Guo,
  • Dongpeng Li,
  • Chaofeng Lan,
  • Chuang Han

摘要

With the continuous development of object detection technology and drone aerial photography technology, object detection technology is increasingly applied in the process of detecting targets captured by drones. However, in the perspective of drones, there will be many small targets, occlusions, and a large number of complex environments. Therefore, this paper proposes the CSF-YOLOv8 small target detection algorithm. The algorithm first proposed CSPPC to replace the C2f in the backbone of YOLOv8, in order to enhance the model’s adaptability to scale changes and its ability to detect small-scale objects. SENetV2 has been added to the output of the small object detection layer in the neck network. This design aims to improve the precision of feature expression and the ability to integrate global information through a multi branch structure, enhancing the algorithm’s detection ability in complex situations and facing targets of different scales. The use of FRMHead detection head has replaced the original detection head, improving the accuracy of object detection in the model. Modify the original loss function CIoU to InnerWIoU, which introduces auxiliary bounding boxes of different scales to calculate the loss. This loss function is suitable for detecting targets with uneven distribution or large size differences. Experimental data shows that the improved algorithm has an 8.4% increase in mAP50 compared to the original algorithm on the VisDrone2019 dataset, mAP50%: 95% increased by 5.9%.