Scholarlink

<p>This study systematically investigates width and depth scaling strategies for YOLOv8n in blood cell detection using the BCCD dataset with 5-fold cross-validation. Contrary to conventional expectations, results show that neither width expansion nor depth increase improves upon the baseline YOLOv8n, which achieves a mean mAP@0.5 of 0.9310 ± 0.0020. Width expansion (0.35<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation>) yields 0.9183 ± 0.0058 (1.4% lower), while depth increase degrades performance further to 0.9116 ± 0.0068 despite 31% more parameters. Gradient analysis reveals that deeper networks retain only 2% of gradient information versus 53% in the baseline, explaining the failure of depth scaling. Fine-grained width analysis confirms an inverse relationship between width and accuracy, with the narrowest 35% configuration outperforming all wider variants. These findings demonstrate that the baseline YOLOv8n already represents the optimal trade-off between accuracy and efficiency for blood cell detection, providing practical guidelines for deploying resource-efficient models in clinical settings.</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

Comparative Analysis of Width and Depth Scaling Strategies for YOLOv8 in Blood Cell Detection

  • 期刊论文

  • 发表日期：

    2026年5月15日

查看全文
  • Jing Yang,
  • Zhenqing Li,
  • Yoshinori Yamaguchi

摘要

This study systematically investigates width and depth scaling strategies for YOLOv8n in blood cell detection using the BCCD dataset with 5-fold cross-validation. Contrary to conventional expectations, results show that neither width expansion nor depth increase improves upon the baseline YOLOv8n, which achieves a mean mAP@0.5 of 0.9310 ± 0.0020. Width expansion (0.35 \(\times \) × ) yields 0.9183 ± 0.0058 (1.4% lower), while depth increase degrades performance further to 0.9116 ± 0.0068 despite 31% more parameters. Gradient analysis reveals that deeper networks retain only 2% of gradient information versus 53% in the baseline, explaining the failure of depth scaling. Fine-grained width analysis confirms an inverse relationship between width and accuracy, with the narrowest 35% configuration outperforming all wider variants. These findings demonstrate that the baseline YOLOv8n already represents the optimal trade-off between accuracy and efficiency for blood cell detection, providing practical guidelines for deploying resource-efficient models in clinical settings.