基于图结构引导的稻穗骨架解析与关键表型参数无损测量方法

Graph Structure-guided Rice Panicle Skeleton Parsing and Non-destructive Measurement of Key Phenotypic Parameters

摘要:

穗部表型参数的高通量、无损获取是水稻育种与表型组学研究的关键环节。针对传统人工测量方法效率低、破坏性强，以及现有图像法依赖人工先验、灵活性差等问题，提出一种基于图结构引导的稻穗骨架解析与关键表型参数无损测量方法。首先，在YOLO v9模型基础上，引入混合背景数据增广与WIoU损失，训练出鲁棒性更强的穗节与穗颈节关键点检测模型；其次，对稻穗图像进行阈值分割与细化，提取其骨架并构建无向图拓扑结构；最后，将模型检测到的关键点与骨架拓扑图深度融合，判别关键点类别，并辅助图论算法自动识别提取穗轴、一次枝梗与二次枝梗，依据标定物实现像素尺度至物理尺度的转换。试验结果表明，优化后的关键点检测模型在穗颈节与穗节检测上的mAP较基准模型分别提高4.5、2.4个百分点，召回率分别提升7.8、4.0个百分点，关键点正确检测比例分别提升4.6、5.0个百分点。在结构计数方面，穗节点计数实现零误差，一次枝梗与二次枝梗计数的平均相对误差分别不超过0.39%、2.38%。在尺度参数测量中，稻穗的一次枝梗、二次枝梗、穗轴长度及穗节间长度的平均相对误差可控制在3.2%、7.5%、3.1%与5.2%以内，平均绝对误差分别不超过2.9、2.3、2.3、1.6mm。本研究实现了稻穗关键表型参数的自动无损提取，可为稻穗表型分析提供一种技术方案。

Abstract:

The high-throughput and non-destructive acquisition of panicle phenotypic parameters is a key step in rice breeding and phenomics research. To address the limitations of traditional manual methods—which are inefficient and destructive—and the poor flexibility of existing image-based methods that rely on manual priors, a graph structure-guided approach for skeleton parsing and non-destructive measurement of key phenotypic parameters was proposed. Firstly, building on the YOLO v9 framework, a more robust key point detection model was trained by incorporating mixed-background data augmentation and a Wise-IoU (WIoU) loss function to improve the detection of panicle nodes and neck nodes. Next, threshold segmentation and thinning were applied to panicle images to extract skeletal structures and construct an undirected graph topology. The detected key points were then deeply integrated with the skeleton topology to classify key point types and assist graph-based algorithms in automatically identifying the rachis, primary branches, and secondary branches. Physical scale conversion was achieved by using a calibration object. Experimental results demonstrated that the improved detection model increased mAP for neck nodes and panicle nodes by 4.5 and 2.4 percentage points, respectively, compared with the baseline, while recall was improved by 7.8 and 4.0 percentage points, and the correct detection ratio of key points was increased by 4.6 and 5.0 percentage points. In structural counting, panicle node counting achieved zero error, and the mean relative errors for primary and secondary branch counts were within 0.39% and 2.38%, respectively. For dimensional measurements, the mean relative errors of primary branch length, secondary branch length, rachis length, and internode length were controlled within 3.2%, 7.5%, 3.1%, and 5.2%, respectively, with mean absolute errors not exceeding 2.9mm, 2.3mm, 2.3mm, and 1.6mm. The research result achieved automatic and non-destructive extraction of key rice panicle phenotypic parameters, providing a viable technical solution for high-throughput panicle phenotyping.