Authors

Abstract

Aiming at the problem of low accuracy of wheat kernel segmentation and counting in wheat breeding tests, this study proposed an improved High-Resolution Network model by embedding coordinate attention mechanism, pyramid pooling module and Lovasz-Softmax loss function. A wheat kernel segmentation dataset with pixel-level labeling was constructed, and systematic comparative experiments were carried out to verify the model performance. The proposed model achieved a mean Intersection over Union of 88.3% on the test set, and the coefficient of determination of kernel counting reached 0.9135. This method can provide technical support for high-throughput phenotyping detection of wheat.

Abstract in Chinese

针对小麦育种试验中穗粒分割计数精度不足的问题，本研究提出一种改进的高分辨率网络模型，通过嵌入坐标注意力机制、金字塔池化模块与 Lovasz-Softmax 损失函数实现模型优化。构建了像素级标注的小麦穗粒分割数据集，通过系统的对比试验验证模型性能。所提模型在测试集上平均交并比达 88.3%，穗粒计数决定系数达 0.9135，可为小麦高通量表型检测提供技术支撑。