Authors

Abstract

Radicle length is an important phenotypic trait for evaluating the growth-promoting effects of plant growth-promoting rhizobacteria on cucumber seeds. However, conventional manual measurement is time-consuming, labor-intensive, and prone to subjective errors, while existing radicle phenotyping systems still suffer from insufficient throughput. To address these limitations, this study proposes an automatic radicle length estimation method for germinated cucumber seeds based on GermHRNet. First, a YOLOv8-p2s model was used to automatically detect seed instances in Petri dish images. Subsequently, a GermHRNet keypoint detection network was developed to accurately localize radicle keypoints. A high-resolution cross-scale feature enhancement module and a dynamic coordinate feature enhancement module were introduced to improve feature representation and keypoint localization accuracy for slender and curved radicles. Finally, radicle length was estimated using a polyline-based pixel distance accumulation strategy according to the detected keypoint coordinates. Experimental results showed that the proposed method achieved an average precision, average recall, and percentage of correct keypoints of 90.6%, 92.2%, and 88.2%, respectively. The mean absolute error, root mean square error, and coefficient of determination for radicle length estimation were 0.51 mm, 0.61 mm, and 0.98, respectively. Under a CPU-only environment without GPU acceleration, the proposed method achieved a throughput of 3.38 × 10⁴ seeds/h. The proposed method enables rapid and accurate radicle length acquisition without specialized equipment and provides an efficient technical solution for large-scale microbial strain screening and seed phenotyping.

Abstract in Chinese

黄瓜种子胚根长度是评价植物促生根际菌促生效应的重要表型参数，但传统人工测量费时费力且主观性强，而现有胚根表型采集系统也存在通量不足的问题。为此，本研究提出一种基于 GermHRNet 的发芽黄瓜种子胚根长度自动估测方法。首先，利用YOLOv8-p2s模型实现培养皿中种子个体的自动检测。随后，构建GermHRNet关键点检测网络，实现胚根关键点的准确定位。通过引入高分辨率交叉特征增强模块和动态坐标特征增强模块，提高了细长弯曲胚根的特征表达与关键点定位精度。最后，基于关键点坐标采用多段折线像素距离累积方法实现胚根长度估测。结果表明：关键点检测的平均精度、平均召回率和正确关键点百分比分别达到90.6%、92.2%和88.2%；长度估测的平均绝对误差、均方根误差和决定系数分别为0.51 mm、0.61 mm和0.98；在普通设备CPU推理且未启用GPU加速的环境下，单机处理通量可达3.38×104粒/h。该方法无需专业设备即可实现发芽黄瓜种子胚根长度的准确、快速获取，可为大规模菌株筛选及种子表型分析提供高效技术支撑。