基于语义分割的复杂背景下黄瓜叶部病害严重程度分级研究  

Research on cucumber leaf disease severity classification in complex background based on semantic segmentation

作　　者：杜鹏飞 黄媛[1,2,3] 高欣娜 武猛[1,2,3] 杜亚茹 杨英茹 Du Pengfei;Huang Yuan;Gao Xinna;Wu Meng;Du Yaru;Yang Yingru(Shijiazhuang Academy of Agricultural and Forestry Sciences,Shijiazhuang,050041,China;Shijiazhuang Agricultural Informational Engineering Technology Innovation Center,Shijiazhuang,050041,China;Hebei Urban Agricultural Technology Innovation Center,Shijiazhuang,050041,China)

机构地区：[1]石家庄市农林科学研究院,石家庄市050041 [2]石家庄市农业信息化工程技术创新中心,石家庄市050041 [3]河北省都市农业技术创新中心,石家庄市050041

出　　处：《中国农机化学报》2023年第11期138-147,共10页Journal of Chinese Agricultural Mechanization

基　　金：河北省重点研发计划(21327410D、21327408D、22327401D);石家庄市农业科技项目(23007)。

摘　　要：为满足实际种植环境下对病害叶片精准用药的需求,以设施环境复杂背景图像为研究对象,提出基于语义分割的复杂背景下黄瓜叶部病害分级方法。首先,使用Labelme标注软件对图像叶片和病斑进行标注,并对部分病害叶片进行图像增强以丰富数据集;然后,改进U-Net网络结构并构建基于深度学习的复杂背景下黄瓜叶片病害分割两阶段架构,对复杂背景下的黄瓜叶片、病斑进行分割;最后,提出黄瓜霜霉病、炭疽病病害严重程度分级模型D-MUNet,对病害等级进行划分。改进后的U-Net模型像素精度、平均交并比和Dice系数分别为90.48%、92.46%、0.645 7,较原始模型提升2.36%、2.34%和0.023 8。黄瓜霜霉病、炭疽病病害分级准确率分别达到92.11%和89.17%。基于语义分割的复杂背景下黄瓜叶部病害严重程度分级方法,能够对黄瓜病害实现有效地分割、分级,为病害的精准防治提供技术支撑。In order to meet the demand for precise medication of diseased leaves in the actual planting environment,a complex background image of facility environment was taken as the research object,and a classification method of cucumber leaf disease under complex background based on semantic segmentation was proposed.First,Labelme labeling software was used to label the image leaves and disease spots,and image enhancement was performed on some diseased leaves to enrich the dataset;then,the U-Net network structure was improved and a deep learning-based cucumber leaf disease segmentation under complex background was constructed.A two-stage architecture is used to segment cucumber leaves and disease spots under complex backgrounds.Finally,the disease severity classification model D-MUNet of cucumber downy mildew and anthracnose is proposed to classify the disease levels.The pixel accuracy,average intersection ratio and Dice coefficient of the improved U-Net model are 90.48%,92.46%,and 0.6457,respectively,which are 2.36%,2.34%,and 0.0238 higher than the original model.The classification accuracy of cucumber downy mildew and anthracnose reaches 92.11%and 89.17%,respectively.The classification method of cucumber leaf disease severity based on semantic segmentation can achieve effective segmentation and classification of cucumber disease,and provide technical support for accurate disease prevention and control.

关 键 词：黄瓜病害 复杂背景 语义分割 两阶段框架 病害严重程度分级 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] S436.421[自动化与计算机技术—计算机科学与技术]