基于语义分割与实例分割的玉米茎秆截面参数测量方法  

Measurement of Maize Stem Cross Section Parameters Based on Semantic Segmentation and Instance Segmentation

作　　者：陈燕[1,2] 李想 曹勉 胡小春 王令强 CHEN Yan;LI Xiang;CAO Mian;HU Xiaochun;WANG Lingqiang(College of Computer and Electronic Information,Guangxi University,Nanning 530004,China;Guangxi Key Laboratory of Multimedia Communications Network Technology,Nanning 530004,China;School of Bigdata and Artificial Intelligence,Guangxi University of Finance and Economics,Nanning 530007,China;College of Agriculture,Guangxi University,Nanning 530004,China)

机构地区：[1]广西大学计算机与电子信息学院,南宁530004 [2]广西多媒体通信与网络技术重点实验室,南宁530004 [3]广西财经学院大数据与人工智能学院,南宁530007 [4]广西大学农学院,南宁530004

出　　处：《农业机械学报》2023年第6期214-222,共9页Transactions of the Chinese Society for Agricultural Machinery

基　　金：广西科学研究与技术开发计划项目(桂科AA20302002-3);广西自然科学基金项目(2020GXNSFAA159090)。

摘　　要：茎秆微观结构与其力学性能密切相关,影响作物的抗倒伏性能。但作物茎秆微观表型参数难以通过人工方式获取,因此急需自动化的测量方法。本研究以玉米为材料,通过光学显微镜获得玉米茎秆横截面切片图像,基于深度学习架构融合ResNet和Unet构建语义分割Res-Unet网络模型,对截面表皮、周皮和髓区3个功能区域进行分割;针对维管束数量多、面积小、密度大的特点,以EfficientDet作为基础网络架构,根据维管束尺寸小的特性,减少双向特征图金字塔(BiFPN)的层数,达到提高推理速度、减少显存占用量的目的,同时添加掩膜分割分支,构造新的网络Eiff-BiFPN实现对维管束的分割。实验结果表明,功能区域分割的平均DICE达到88.17%;维管束分割的AP50和AP50:70分别达到88.78%和72.80%。根据分割结果,可以获得玉米茎秆截面尺寸、各功能区域尺寸和维管束数量、面积等微观结构参数。本文方法具有精确性、实时性和可用性,可用于玉米茎秆微观结构参数的自动化测定,为作物抗倒伏研究提供技术基础。Stem microstructure is closely related to its mechanical properties and affects lodging resistance in crops.But crop microphenotypic parameters are difficult to obtain manually.Therefore,automated measurement methods are urgently needed.The lack of measurement methods for high⁃throughput vascular bundle parameters seriously restricts the in⁃depth study.Based on the deep learning architecture,ResNet and Unet network were merged to construct the semantic segmentation model Res Unet to segment function zones in maize stem cross section.In view of the small area,large number and dense distribution of vascular bundle in maize stem cross section,EfficientDet was used as the basic network architecture.According to the characteristics of small size of vascular bundles,the number of layers of BiFPN was reduced to improve reasoning speed and reduce the occupation of video memory.Mask segmentation branches were added to construct a new network Eiff BiFPN to segment vascular bundles.The results showed that the DICE of each function zone could reach an average DICE of 88.17%,and the vascular bundle segmentation task could reach 88.78%and 72.80%on AP50 and AP50:70,respectively.Therefore,the proposed method was accurate,real⁃time and available,which can be used for automatic determination of microstructure parameters of maize stem,and a technical basis was established for the study of crop lodging resistance.According to the segmentation results,the crosssectional size of corn stem,the size of each functional area,the number and area of vascular bundles and other microstructure parameters can be obtained.

关 键 词：玉米茎秆 截面参数 语义分割 实例分割 显微结构 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]