Abstract:

The precisive segmentation of grapevine structure is an important prerequisite for reasoning and locating the pruning points. To precisely segmentate grapevine structure， this article established a vine structure data set under natural planting conditions， and proposed a grape vine structure division method based on the U-net model. Through the comparative experiment of backbone feature extraction network and model segmentation performance， the optimal U-net model structure was obtained and its segmentation performance under different density degree targets was verified. The results showed that the precision of the U-net model with VGG 16 as the backbone feature extraction network was 93.55%， the recall was 94.15%， the mean pixel accuracy was 94.15%， and the mean intersection over union was 88.65%. Compared with traditional image segmation methods and control group model segmentation effects， it could ensure that the structure of the grape vines was complete in the context of natural planting， and the connection relationship between the structure was correct， so it was suitable for the segmentation task of grapevine structures with shade between plants， laid the foundation for achieving intelligent grape vines in winter pruning operations.

Key words: machine vision, image processing, grapevine structure, deep learning

摘要：

葡萄藤结构的精确分割是推理与定位冬季剪枝点位置的重要前提。为精确分割葡萄藤结构，建立自然种植条件下的葡萄藤结构数据集，提出一种基于U-net模型的葡萄藤结构分割方法，通过主干特征提取网络和模型分割性能的对比试验，得出最优的U-net模型结构并验证其在不同疏密程度目标下的分割性能。结果表明，以VGG 16为主干特征提取网络的U-net模型准确率达93.55%、召回率为94.15%、类别平均像素准确率为94.15%、均交并比为88.65%，与传统图像分割方法和对照组模型分割效果相比，其能保证自然种植背景下葡萄藤各结构分割边缘完整，结构之间连接关系正确，可适用于植株间存在遮挡的葡萄藤结构分割任务，为实现智能化葡萄藤冬季剪枝作业奠定基础。

关键词: 机器视觉, 图像分割, 葡萄藤结构, 深度学习