Abstract: The precise identification and positioning of pruning points is the basis for the intelligent pruning of grapevines in winter, the segmentation of the critical structure of the grapevine is an important prerequisite for reasoning about the precise pruning point. Aiming at the problem that the existing cutting method is greatly affected by the background, resulting in the loss of critical structures of the grapevine, and inaccurate identification and positioning of pruning points, a segmentation method of grapevine critical structure based on Mask R-CNN was proposed, the grapevine pruning model and the critical structure data sets were established. Through the comparative experiment of backbone feature extraction network and segmentation performance, the optimal Mask R-CNN model structure was obtained and its fitting and generalization ability and segmentation performance in different natural backgrounds were verified, The results showed that the Mask R-CNN model with ResNet 101+FPN as the backbone feature extraction network proposed had better fitting and generalization ability, compared with the control group model, the accuracy rate was increased by 7.33% and 8.89%, the recall rate was increased by 9.32% and 9.26%, and the average precision was increased by 12.69% and 12.63% respectively, it could overcome various natural planting background factors, the edge of the segmentation target was complete, and the connection relationship between the critical structures of the grapevine was correct.

Key words: image processing, feature extraction, grapevine, deep learning

摘要： 剪枝点的精确识别与定位是实现葡萄藤冬季剪枝智能化的基础,葡萄藤关键结构的分割是用于推理精确剪枝点的重要前提。针对现有分割方法受背景影响较大致使葡萄藤各关键结构损失和剪枝点识别与定位不准确的问题,提出一种基于Mask R-CNN的葡萄藤关键结构分割方法,建立葡萄藤修剪模型以及各关键结构数据集。通过主干特征提取网络和分割性能的对比试验,得出最优的Mask R-CNN模型结构并验证其拟合与泛化能力以及在不同自然背景下的分割性能。结果表明,以ResNet 101+FPN为主干特征提取网络的Mask R-CNN模型具有较好的拟合与泛化能力,相较于对照组模型准确率分别提升7.33%和8.89%,召回率分别提升9.32%和9.26%,平均精度均值分别提升12.69%和12.63%,其能够克服各类自然种植背景因素,分割目标边缘完整,葡萄藤各关键结构之间连接关系正确。

关键词: 图像分割, 特征提取, 葡萄藤, 深度学习