葡萄多模态目标检测和语义分割数据集  

A dataset of grape multimodal object detection and semantic segmentation

作　　者：陈文骏 饶元[1,2,3] 王丰仪 张雨 杨雨梦 罗庆 张通 万天与 刘心宇 张梦宇 张蕊 CHEN Wenjun;RAO Yuan;WANG Fengyi;ZHANG Yu;YANG Yumeng;LUO Qing;ZHANG Tong;WAN Tianyu;LIU Xinyu;ZHANG Mengyu;ZHANG Rui(College of Information and Computer Science,Anhui Agricultural University,Hefei 230036,P.R.China;Key Laboratory of Agricultural Sensors,Ministry of Agriculture and Rural Affairs,Hefei 230036,P.R.China;Anhui Provincial Key Laboratory of Smart Agricultural Technology and Equipment,Hefei 230036,P.R.China)

机构地区：[1]安徽农业大学,信息与计算机学院,合肥230036 [2]农业农村部农业传感器重点实验室,合肥230036 [3]智慧农业技术与装备安徽省重点实验室,合肥230036

出　　处：《中国科学数据(中英文网络版)》2025年第1期82-97,共16页China Scientific Data

基　　金：国家自然科学基金(32371993);安徽省重点研究和开发计划面上攻关项目(202104a06020012,202204c06020022);安徽省高校自然科学重大项目(2022AH040125)。

摘　　要：葡萄果实的采摘点定位准确率依赖于目标检测和语义分割网络的性能。然而,在实际应用场景中,基于可见光图像的葡萄果实目标识别准确率和分割精度易受光照变化、复杂环境影响,往往表现不佳,且葡萄果实成串生长,现有苹果、梨子等多模态数据集难以满足串形葡萄果实的识别需求。构建基于可见光、深度、近红外的葡萄多模态目标检测和语义分割数据集,对于探索更高识别率和更强泛化能力的葡萄果实目标检测和语义分割模型至关重要。本数据集约39.08 GB,共收集了在不同光照和遮挡条件下青色、紫色两类6个品种的葡萄高质量多模态视频流数据,并从中提取3954张图像样本进行语义分割和目标检测标注。在使用旋转、缩放、错切、平移,以及高斯模糊等图像增强手段下,可满足主流深度学习模型训练需要。本数据集可为多模态视觉数据融合、葡萄果实语义分割和目标检测等领域提供宝贵的基础数据资源,对促进农机装备智能化领域研究具有重要的实际应用价值。The accuracy of grape picking point localization relies on the performance of grape detection and semantic segmentation network.However,in practical application scenarios,the accuracy and segmentation precision of grape targets based on visible light images are susceptible to light variations and complex environments,often resulting in poor performance.Moreover,grapes grow in bunches,and the existing multimodal datasets for apples and pears may not adequately fulfill the recognition needs of bunch-shaped grapes.The construction of a dataset of grape multimodal object detection and semantic segmentation based on visible light,depth,and near-infrared multimodal object detection and semantic segmentation is crucial to exploring better recognition rates and stronger generalization capabilities for grape detection and semantic segmentation models.This dataset,totaling about 39.08 GB,contains high-quality multimodal video stream data of green and purple grapes,including six varieties captured under different illumination and obscuration conditions.Additionally,the dataset offers 3,954 labeled image samples extracted from the aforementioned multimodal video stream.Using techniques such as rotation,deflation,mis-slicing,panning,and Gaussian blur,the dataset can be augmented for the training implementation of mainstream deep learning models.The dataset can provide valuable basic data resources for multimodal fusion,grape semantic segmentation,and object detection,and have important practical application value for promoting research in the field of agricultural machinery and equipment intelligence.

关 键 词：自然环境 葡萄 多模态 语义分割 目标检测 

分 类 号：S663.1[农业科学—果树学] TP391.41[农业科学—园艺学]