Authors

Abstract

In greenhouse cultivation of cherry tomatoes, the accurate identification of pollination status is a key prerequisite for achieving automated robotic pollination. As the site of direct interaction during pollination, changes in the colour and morphology of the stigma most accurately indicate whether pollination has been successful. However, due to the minuscule size of the stigma and the extremely subtle visual differences among the three states—unpollinated, pollinated, and fruiting—combined with interference from fluctuating lighting conditions and complex backgrounds in greenhouse environments, traditional methods struggle to provide high-precision visual perception data for pollination robots. To address this issue, this paper proposes an improved model, YOLO11-RCSP, which focuses on the stigma as the core recognition object to achieve automatic classification of the three pollination states: introducing a novel attention mechanism, RFA (Receptive Field Attention), adding an enhancement module (CPA-Enhancer) to the original model, and incorporating SAConv switchable dilated convolutions. In addition, the loss function is improved using the Powerful-IoU method (adaptive penalty factor and gradient adjustment function based on anchor box quality). The precision rate of the improved model reached 88.40%, the recall rate reached 84.60%, and the average precision was 90.80% when the IoU was 0.50. The average precision ranged from 0.50 to 0.95, reaching 63.20%. These metrics represent improvements of 5.20%, 3.30%, 4.40%, and 5.30%, respectively, compared to the original YOLO11 network model, demonstrating stronger feature extraction capabilities and robustness. Ablation experiments further confirm that the attention mechanism and multi-scale optimization have a synergistic effect on enhancing the performance of stigma pollination status recognition. This research will be applied to pistil-level pollination status recognition tasks, helping to advance the development and application of pollination robots, and providing critical technical support for the development of automated pollination equipment in protected agriculture.

Abstract in Chinese

在圣女果温室栽培中，授粉状态的精准识别是实现机器人自动化授粉作业的核心前提。柱头作为授粉过程的直接作用部位，其颜色、形态变化能够最准确地反映授粉是否成功。然而柱头目标尺度微小，未授粉、授粉和结果三类状态的视觉差异极为细微，而且受温室环境中光照变化与复杂背景的干扰，传统方法难以为授粉机器人提供高精度的视觉感知决策依据。针对该问题，本文提出一个改进模型YOLO11-RCSP，以柱头为核心识别对象，实现三类授粉状态的自动分类: 引入一种新型的注意力机制RFA（感受野注意力）；在原有模型中添加一种增强模块（CPA-Enhancer）；添加SAConv可切换空洞卷积；改进损失函数采用Powerful-IoU(自适应惩罚因子和基于锚框质量的梯度调节函数) 的方法。改进后模型的准确率达到88.40%，召回率达到84.60%，平均精度在IoU=0.50时为90.80%，在0.50至0.95的IoU平均精度达到63.20%，各项指标相较于原YOLO11网络模型分别提高了5.20%，3.30%，4.40%，5.30%，拥有更强的特征提取能力和鲁棒性。消融实验进一步证实注意力机制与多尺度优化对提升柱头授粉状态识别性能具有协同效应。本研究将应用于柱头级别的授粉状态识别任务，有助于推动授粉机器人的研发与应用，为设施农业自动化授粉装备的研发提供关键技术支撑。