Authors

Abstract

Aiming at the low accuracy of wheat kernel segmentation and counting caused by severe adhesion, blurred boundaries and complex field environment in wheat breeding tests, this study proposed an improved High-Resolution Network (HRNet) model optimized by integrating coordinate attention (CA), pyramid pooling module (PPM) and Lovasz-Softmax loss function. A pixel-level labeled wheat kernel segmentation dataset (3600 images, 5 national standard spike types) was constructed based on field samples from the Huang-Huai-Hai winter wheat region. The proposed model achieved a mean Intersection over Union (mIoU) of 88.3% on the laboratory test set, and a counting determination coefficient (R²) of 0.9135. For agricultural engineering application, the trained model was imported into a breeding phenotype analysis workstation, forming a complete application process of "field sampling → standardized image acquisition → batch model inference → automatic counting → phenotype data output". Field verification showed that the model realized high-throughput counting with an efficiency of 500 spikes per hour (38 times higher than manual counting), and maintained stable accuracy in complex field environments. This method can provide efficient technical support for high-throughput wheat phenotyping detection and precision agriculture.

Abstract in Chinese

针对小麦育种试验中穗粒粘连严重、边界模糊、田间环境复杂导致的穗粒分割计数精度不足的问题，本研究融合坐标注意力机制（CA）、金字塔池化模块（PPM）与 Lovasz-Softmax 损失函数，提出一种改进的高分辨率网络（HRNet）模型。基于黄淮海冬麦区田间样本，构建了包含 3600 张图像、覆盖 5 种国标穗型的像素级标注小麦穗粒分割数据集。所提模型在实验室测试集上平均交并比（mIoU）达 88.3%，穗粒计数决定系数（R²）达 0.9135。面向农业工程应用，将训练完成的模型导入育种表型分析工作站，形成了 “田间采样→标准化图像采集→批量模型推理→自动计数→表型数据输出” 的完整应用流程。田间验证表明，该模型可实现高通量穗粒计数，作业效率达 500 穗 / 小时（较人工计数提升 38 倍），在复杂田间环境下精度保持稳定，可为小麦高通量表型检测与精准农业生产提供高效技术支撑。