Scholarlink

<p>This paper proposes a lightweight detection algorithm based on progressive feature fusion specifically for the scenario of dense small target detection, aiming to optimize the detection performance for small targets. Firstly, addressing the inadequate extraction of small target feature information by the C2f module in YOLOv8n, the Context Anchor Attention (CAA) mechanism is introduced to enhance the feature representation of small targets. Secondly, to balance the potential increase in parameter count caused by introducing CAA, depthwise convolution technology is incorporated. Additionally, to provide more detailed information about small targets, the high-resolution feature map at the P2 level is added as an input to the feature fusion component. Finally, addressing the information conflicts that may arise from cross-scale fusion in the Path Aggregation Feature Pyramid Network (PAFPN), this paper proposes a novel feature fusion network based on the Asymptotic Feature Pyramid Network (AFPN). Compared to YOLOv8n, the improved model achieves a 27.9% reduction in the number of parameters, a 2.4% increase in mAP@0.5, and a 1.5% increase in mAP@0.5:0.95 on the Visdrone2019 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

GDA-YOLOv8n: a lightweight small object detection algorithm based on gradual feature aggregation

  • 期刊论文

  • 发表日期：

    2025年7月30日

查看全文
  • Xiaolin Li,
  • Pengjiang Wang,
  • Jiahang Wang,
  • Fanqi Zeng,
  • Xu Liu

摘要

This paper proposes a lightweight detection algorithm based on progressive feature fusion specifically for the scenario of dense small target detection, aiming to optimize the detection performance for small targets. Firstly, addressing the inadequate extraction of small target feature information by the C2f module in YOLOv8n, the Context Anchor Attention (CAA) mechanism is introduced to enhance the feature representation of small targets. Secondly, to balance the potential increase in parameter count caused by introducing CAA, depthwise convolution technology is incorporated. Additionally, to provide more detailed information about small targets, the high-resolution feature map at the P2 level is added as an input to the feature fusion component. Finally, addressing the information conflicts that may arise from cross-scale fusion in the Path Aggregation Feature Pyramid Network (PAFPN), this paper proposes a novel feature fusion network based on the Asymptotic Feature Pyramid Network (AFPN). Compared to YOLOv8n, the improved model achieves a 27.9% reduction in the number of parameters, a 2.4% increase in mAP@0.5, and a 1.5% increase in mAP@0.5:0.95 on the Visdrone2019 dataset.