适应遮挡条件下奶油生菜的实例分割方法研究  

Research on Instance Segmentation Method of Butter Lettuce Under Adaptive Occlusion Conditions

作　　者：韩江枫 杨意 郑鸿燊 刘厚诚[3] 琚俊 辜松[1] Han Jiangfeng;Yang Yi;Zheng Hongshen;Liu Houcheng;Ju Jun;Gu Song(College of Engineering,South China Agricultural University,Guangzhou 510642,China;College of Electronic Engineering,South China Agricultural University,Guangzhou 510642,China;College of Horticulture,South China Agricultural University,Guangzhou 510642,China)

机构地区：[1]华南农业大学工程学院,广州510642 [2]华南农业大学电子工程学院,广州510642 [3]华南农业大学园艺学院,广州510642

出　　处：《农机化研究》2024年第8期80-84,共5页Journal of Agricultural Mechanization Research

基　　金：国家重点研发计划项目(2021YFD2000703);广东省现代农业产业共性关键技术研发创新团队建设项目(2022KJ131)。

摘　　要：利用机器视觉技术测量生菜的表型参数对探索生菜的生长规律有着非常重要的意义,而构建生菜个体的识别及轮廓分割算法是实现表型参数精准测量的重要前提;但是,在生菜培育至采收期,俯视图下生菜个体间叶片相互重叠遮挡,对个体识别和轮廓分割造成很大的阻碍。为此,改进了Mask R-CNN神经网络模型,掩膜分支采用class-agnostic模式,以ResNeXt50联合FPN替换原有的卷积主干,实现了遮挡条件下奶油生菜的个体识别和轮廓分割。为了对改进模型的分割精度进行验证分析,采用平均精度AP75和平均检测耗时作为评价指标,与原始Mask R-CNN、DeepMask、MNC分割模型在不同程度遮挡测试集上设置对比试验。结果表明:改进模型的平均精度达到98.7%,相比原模型提高了约4%,且在重度遮挡测试集上依然能够保持良好的分割精度。研究结果可为遮挡条件下植物叶片的识别和分割提供算法参考,也可为奶油生菜的表型参数提取提供技术支持。Using machine vision technology to measure the phenotypic parameters of lettuce is of great significance to ex-plore the growth law of lettuce.The construction of lettuce individual identification and outer contour segmentation algo-rithms is an important prerequisite for accurate measurement of phenotypic parameters,but when lettuce is cultivated to harvest In the top view,the leaves of the lettuce individuals overlap and block each other,which greatly hinders the indi-vidual identification and outer contour segmentation of lettuce.In response to the above problems,this paper improves the Mask R-CNN neural network model,the mask branch adopts the class-agnostic mode,and the original convolution backbone is replaced by ResNeXt50 combined with FPN,which realizes the individual recognition and outer contour seg-mentation of butter lettuce under occlusion conditions.In order to verify and analyze the segmentation accuracy of the im-proved model,this paper uses the average accuracy AP75 and the average detection time as the evaluation indicators,and sets up comparative experiments with the original Mask R-CNN,DeepMask,and MNC segmentation models on dif-ferent degrees of occlusion test sets.The results show that the average accuracy of the improved model reaches 98.7%,which is about 4%higher than the original model,and it can still maintain good segmentation accuracy on the heavily oc-cluded test set.This study can provide an algorithm reference for the identification and segmentation of plant leaves under shading conditions,and also provide technical support for the extraction of phenotypic parameters of butter lettuce.

关 键 词：奶油生菜 轮廓分割 遮挡 Mask R-CNN 深度学习 图像处理 

分 类 号：S126[农业科学—农业基础科学] TP391.41[自动化与计算机技术—计算机应用技术]