Abstract: Targeting at the problem of automatic acquisition of stem node number and stem thickness phenotypes in soybean plants，this study decomposed the task into several sub⁃problems including stem node identification,stem region detection,stem node localization and sorting,and stem thickness calculation,and proposed a method of automatic detection of stem⁃related phenotypes that integrated target detection,semantic segmentation,and path planning.Based on A^* path planning algorithm and target detection algorithm,the automatic detection and count of stem node were realized.Based on analogic method,curvature method and the maximum tangent circle method,automatic measurement and calculation of stem diameter were realized.By incoporating the coordinate attention mechanism into the YOLOv5s baseline model，the model achieved an average detection accuracy of 93.2% for stem node identification,which was also optimized by fusing a filtering algorithm incorporating the stem node position information to address the adverse effect of pod overlapping in dense plants at the maturity stage,which resulted in a 2.4 percentage point improvement when compared to the original YOLOv5s.At the same time,the mean intersection over union of U2⁃Net semantic segmentation network for segmentation of stems reached 82.6% after optimizing. Based on the stem regions after semantic segmentation,compared with the three stem diameter detection methods proposed in this paper,the maximum tangent circle method had the highest accuracy. Results on real plant samples showed that for the number of stem nodes and main stem diameter,the mean absolute errors of the optimized method were 1.33 and 0.99 mm respectively,which were basically consistent with the threshold requirement for manual measurement data.The root mean square errors were 1.74 and 1.20 respectively,and the mean average percentage errors were 8.96% and 16.37% respectively.

Key words: Soybean plant, Stem node, Stem diameter, Phenotype automatic detection, A* algorithm, The maximum tangent circle method

摘要： 针对成熟期大豆植株中茎节数和茎粗表型的自动获取问题，将其分解为茎节识别、茎秆区域检测、茎节定位与排序、茎粗计算等若干子问题，提出一种融合目标检测、语义分割、路径规划等算法的茎秆表型自动检测方法。以A^*寻径算法和目标检测算法为基础，实现了茎节自动检测与计数；以类比法、曲率法和最大内切圆法为基础，实现了茎粗自动检测与计算。针对成熟期植株中豆荚遮挡导致茎秆表型检测精度不佳的问题，在以YOLOv5s为基准的目标检测模型上融入坐标注意力机制，以增强对被遮挡茎节的关注，同时，针对豆荚密集导致的茎节误检问题，设计融合位置信息的过滤算法，将茎节检测任务的平均检测准确度（mAP）提升至93.2%，高于基准模型2.4百分点。使用语义分割模型识别茎秆区域，优化后的均交并比（mIoU）达82.6%。基于语义分割后的茎秆区域，对比3种茎粗检测方法，以最大内切圆法准确率最高。在实际植株样本图片上的检测结果表明，所提出的方法在茎节数和茎粗表型上的平均绝对误差分别为1.33个和0.99 mm，与人工测量的阈值要求基本相符，均方根误差分别为1.74和1.20，平均绝对百分比误差分别为8.96%和16.37%。

关键词: 大豆植株, 茎节, 茎粗, 表型自动检测, A^*算法, 最大内切圆法