Scholarlink

<p>The automated assessment of tomato ripeness is vital for modern greenhouse operations, yet challenges remain due to variable environmental conditions. To provide a solution, we propose rank-aware You Only Look Once (YOLO), a novel detection framework that incorporates the biological prior of top-to-bottom ripening within fruit clusters. This is achieved through two key innovations: an efficient position-aware head for regressing relative height for fruits and a dynamic margin-aware ranking loss (DM-RankLoss) that enforces the correct spatial sequence. Evaluated on a 3500-image dataset from a solar greenhouse, our plug-and-play module could boost the mean average precision (mAP) at intersection over union (IoU) threshold of 0.50 (mAP<sub>50</sub>) of multiple YOLO architectures by up to 5.66 pecentage points. The model effectively learns the cluster topology, achieving a height-mean absolute error (H-MAE) of 0.107 (normalized) and a pairwise ranking accuracy (PRA) of 84.59%, while it reduces the parameter count by over 10% compared to the baseline for efficient deployment. Visualizations confirm that the model leverages spatial context to resolve color ambiguities. Our work offers a sensor-free, accurate, and efficient solution for in situ phenotyping in agricultural robotics.</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

Embedding of ripening topology into one-stage detection for tomato cluster phenotyping

  • 期刊论文

  • 发表日期：

    2026年5月23日

查看全文
  • Bingquan Chu,
  • Ruiyuan Wu,
  • Haijun Zhang,
  • Haochuan Qin,
  • Zishun Peng,
  • Fengle Zhu,
  • Yong He

摘要

The automated assessment of tomato ripeness is vital for modern greenhouse operations, yet challenges remain due to variable environmental conditions. To provide a solution, we propose rank-aware You Only Look Once (YOLO), a novel detection framework that incorporates the biological prior of top-to-bottom ripening within fruit clusters. This is achieved through two key innovations: an efficient position-aware head for regressing relative height for fruits and a dynamic margin-aware ranking loss (DM-RankLoss) that enforces the correct spatial sequence. Evaluated on a 3500-image dataset from a solar greenhouse, our plug-and-play module could boost the mean average precision (mAP) at intersection over union (IoU) threshold of 0.50 (mAP50) of multiple YOLO architectures by up to 5.66 pecentage points. The model effectively learns the cluster topology, achieving a height-mean absolute error (H-MAE) of 0.107 (normalized) and a pairwise ranking accuracy (PRA) of 84.59%, while it reduces the parameter count by over 10% compared to the baseline for efficient deployment. Visualizations confirm that the model leverages spatial context to resolve color ambiguities. Our work offers a sensor-free, accurate, and efficient solution for in situ phenotyping in agricultural robotics.