基于图像处理和聚类算法的待考种大豆主茎节数统计

为了实现待考种大豆植株主茎节数的快速、高效测量，提出一种基于图像处理和聚类算法的待考种大豆主茎节数统计方法。首先，获取不同视角下的已脱叶待考种大豆植株图像，随机抽取训练集与验证集样本植株，并设定初始图像采集间隔与抽样步长；其次，通过植株分割、骨架提取、主茎节点去噪等操作，获取分布于植株主茎上的待检测大豆茎节点；通过基于空间距离的数据转换方法将分布离散的大豆茎节点转换至便于聚类的数据集内；利用HDBSCAN聚类算法对不同采集视角下的待检测大豆茎节点进行聚类，统计、记录主茎节数识别准确率，筛选最优采集间隔；最后，利用最优采集间隔对剩余样本植株主茎节数进行统计、分析。在63株 “中黄30”待考种大豆植株中抽取21株植株作为训练集，并进行实验测试，发现在采集间隔为90°时，以最小聚类簇为2，融合处理4幅大豆图像，大豆主茎节数识别效果最优。据此对42株验证集样本植株进行主茎节数识别和分析，结果表明，大豆主茎节数识别准确率可达98.25%。该方法能够快速、准确获取大豆主茎节数，可满足大豆考种需求。

Statistics of Seed-testing Soybean Main Stem Nodes Based on Image Processing and Clustering Algorithm

Aiming at measuring the number of main stem nodes of soybean plants quickly and efficiently, a statistical method of soybean main stem nodes was proposed based on image processing and clustering algorithm. Firstly, multiple perspectives of the soybean plants were obtained by using a camera, the initial image collection interval and sampling step were set, and then some of the plants were extracted as the train set and validation set. Secondly, through the operations such as plant segmentation, skeleton extraction, and denoising of the main stem nodes, the soybean main stem nodes to be detected were obtained. Meanwhile, the multiple dimensional scaling (MDS) was used to convert the data into a space which was easy to cluster. Then, the hierarchical density based spatial clustering of applications with noise (HDBSCAN) clustering algorithm was used to cluster the soybean stem nodes to be detected from multiple perspectives, and the recognition accuracy of the number of main stem nodes were recorded. Finally, the optimal collection interval was used to determine the number of main stem nodes of the remaining sample plants and conduct statistical analysis. The experiments were carried out based on the above method by using 63 samples which variety was called Zhonghuang 30. 21 plants were selected as the training set, and it turned out that, under the condition of 90° interval, and four soybean images were captured and fused with the minimum cluster of 2, the node number recognition results were mostly distributed in the effective range. To identify and analyze the main stem node number of the remaining 42 sample plants, the corresponding soybean main stem node number recognition accuracy rate can reach 98.25%. The experiment results showed that this method can meet the needs of soybean plant test requirements.