基于改进Mask R-CNN的黄瓜叶面积测量模型  

Measuring the cucumber leaf area using improved Mask R-CNN

作　　者：章永龙[1] 解亚玲 徐向英[1] 缪旻珉[2,3,4] ZHANG Yonglong;XIE Yaling;XU Xiangying;MIAO Minmin(School of Information Engineering,Yangzhou University,Yangzhou 225000,China;School of Horticulture and Plant Protection,Yangzhou University,Yangzhou 225000,China;Joint International Research Laboratory of Agriculture and Agri-Product Safety of Ministry of Education of China,Yangzhou 225000,China;Key Laboratory of Plant Functional Genomics of the Ministry of Education/Jiangsu Key Laboratory of Crop Genomics and Molecular Breeding,Yangzhou 225000,China)

机构地区：[1]扬州大学信息工程学院,扬州225000 [2]扬州大学园艺与植物保护学院,扬州225000 [3]教育部农业与农产品安全国际联合实验室,扬州225000 [4]教育部植物功能基因组学重点实验室/江苏省作物基因组学与分子育种重点实验室,扬州225000

出　　处：《农业工程学报》2023年第17期182-189,共8页Transactions of the Chinese Society of Agricultural Engineering

基　　金：江苏研究与试验发展(R&D)基金(BE2022425);扬州市市校合作专项(YZ2021150,YZ2022179)。

摘　　要：植物叶面积可以反映出植物的生长速率、养分吸收以及光合作用能力,针对锯齿状边缘的黄瓜叶片分割精度较低,叶面积测量误差较大等问题。该研究提出一种深度卷积网络模型Marm,在Mask R-CNN的基础上利用Sobel算子进行边缘检测,使模型生成的掩膜更接近叶片的边缘。另外,引入边缘损失以提升叶片边缘的分割精度。借助参照物标签,利用模型输出的掩膜图像进行面积计算,获得黄瓜叶片在不同生长周期的叶面积。试验结果表明,Marm模型精确率、召回率和交并比达到99.1%、94.87%和92.18%,比原始的Mask R-CNN分别提高1.28个百分点、1.13个百分点和1.05个百分点,面积误差率下降1.43个百分点。当图像中存在叶片遮挡和阴影等多种影响,黄瓜叶片的面积误差率仍然能保持在5.45%左右。该研究有效解决了锯齿状边缘的叶片分割问题,将为植物表型研究提供技术支撑。The plant leaf area can serve as an indicator of the plant's growth rate,nutrient uptake,and photosynthetic efficiency.However,some challenges are still remained to accurately segment the leaf areas and subsequent measurements,particularly with the leaves bearing some distinct characteristics,such as the serrated edges found on cucumber leaves.It is a high demand to promote the reliability of research outcomes and practical applications in the field of agriculture.In this study,an innovative"Marm"model was introduced to measure the cucumber leaf area from the realm of deep convolutional neural networks(CNN).The architecture of Mask R-CNN was also extended to incorporate the Sobel operator.The intricate contours of cucumber leaf edges were capture and detect using the Sobel operator with the enhanced edge details.A suite of mask was obtained to fully depicted the intricate topography inherent in these leaves.Among them,the edges were acted as a connective thread weaving precision into the process.The architecture of the model was enhanced to introduce a tailored component of edge loss.A strategic addition was then used to improve the segmentation accuracies.A weight factor was introduced into the new loss component,in order to balance among the existing loss functions and potentially computational costs.The better performance of the improved model was then effectively enhanced with the pragmatic feasibility.A comprehensive strategy was employed to leverage a synergistic blend of model-generated mask images and validated reference object labels.The morphological intricacies and contextual validation were integrated to facilitated the steadfast and reliable measurements of cucumber leaf area across various growth stages.A better resilience of data and validation was then established to deftly navigate the diverse terrain of leaf morphology from germination to maturity.The precision,recall,and Intersection over Union(IoU)scores reached 99.1%,94.87%,and 92.18%,respectively,after the empirical validation,indicating ove

关 键 词：图像分割 深度学习 Mask R-CNN 植物表型 实例分割 黄瓜 

分 类 号：S24[农业科学—农业电气化与自动化]