Abstract: Grape precise instance segmentation is a key technology for intelligent agricultural management in grape industry parks, and improving the accuracy of grape segmentation is of great significance for the automation of smart vineyards. Due to the dense growth and serious overlapping of grapes, traditional Mask R-CNN algorithms often suffer from inaccurate instance boundaries and errors in overlapping regions. To address these issues, a grape precise segmentation algorithm based on local weighted incoherent region supervision is proposed. Firstly, an additional incoherent region prediction branch is embedded to extract the incoherent regions (edges) and local incoherent regions (overlapping boundaries) of the grapes. Then, the extracted masks of the aforementioned regions are subjected to weighted supervision to enhance the  ability of network to extract features from error-prone areas. Finally, a pixel-affinity-based multitask fusion module is proposed to model the pixel affinity relationship between the incoherent branch and the mask branch, and further improve the network instance segmentation performance. This research method is compared with instance segmentation network models such as YOLACT++, Mask R-CNN, and SOLOv2 using different evaluation metrics. Experimental results demonstrate that this method exhibits good segmentation performance and can accurately segment grape boundaries and instance overlapping areas.

Key words: grape precise segmentation, local incoherent regions, weighted supervision, pixel affinity fusion

摘要： 精确实例分割作为葡萄产业园中智能化农艺管理的一项关键技术，对于智慧葡萄园的自动化管理具有重要意义。由于葡萄生长密集、重叠现象严重，传统Mask R-CNN算法分割葡萄时存在实例边缘不准确且重叠区域易错的问题。针对上述问题，提出一种基于局部加权非相干区域监督的葡萄精确分割算法。嵌入额外的非相干区域预测分支，提取葡萄的非相干区域（边缘）和局部非相干区域（重叠边界）；对于提取到的上述区域掩码进行加权监督，增强网络对易错区域的特征提取能力；基于像素亲和的多任务融合模块被提出，用于建模非相干分支和掩码分支间的像素亲和关系，进一步提升网络分割性能。利用该方法在与YOLACT++、Mask R-CNN、SOLOv2等实例分割网络在不同的评价指标上进行比较，实验结果表明该方法具有较好的分割性能，可以精确分割出葡萄边界及实例重叠区域。

关键词: 葡萄精确分割, 局部非相干区域, 加权监督, 像素亲和性融合