摘要
以识别玉米秧苗茎秆为目标,采用云台搭载电荷耦合器件(CCD)相机获得玉米秧苗图像,采用Label Image插件制作了玉米秧苗的标记与标签。基于深度学习框架Tensor Flow搭建了多尺度分层特征的卷积神经网络模型,应用4倍膨胀的单位卷积核,获得了玉米秧苗图像的识别模型,其识别准确率为99.65%。将已知玉米秧苗图像划分为最佳子块,求取了各个子块的最佳二值化阈值。选取6种杂草密度在每天5个时间段进行为期3 d的试验,共采集了10800幅图像。试验结果显示,对玉米秧苗茎秆的平均识别准确率为98.93%,且光照条件与田间杂草密度对识别结果没有显著影响(P>0.05)。
Modern agricultural equipment is developing towards the intelligent machinery,and deep learning and machine vision are core technologies in realizing the intelligent machinery. In terms of the machine-vision based intelligent agricultural machinery that works in maize field,they should forward towards at the maize stem or avoiding the maize stem,while the ability to identify the maize stem accurately is the premise to ensure them work properly. Aiming to distinguish the maize( Zea Mays L.)stems,which grew in the field. In order to acquire the high quality field pictures,one charge-coupled device( CCD) camera was mounted in one camera gimbal. The plug-in unit named after LabelImage was applied to mark and label the maize plants,based on the deep learning framework TensorFlow,a convolution neural network model with multi-scale hierarchical features was built, and the unit convolution kernel with four times expansion was applied. Thus the maize seedling recognition model was obtained with the recognition accuracy of 99. 65%. Based on the recognition results,the morphological process was conducted by the OpenCV 3. 4. 2 and the Python 3. 6. 5. The threshold value exerted the major influence on depending the information completeness during the binary process,the pictures that contained the maize seedlings were divided into an optimum parts,then an optimum threshold value for each part would be calculated by the algorithm that described. Inspired by the bounding rectangle of each object was different in the binary picture,the aspect ratio was utilized to distinguish the maize seedling stem,and the minimum aspect ratio was computed,then the corresponding bounding rectangle was filled red which indicated the stems. The field experiment was conducted from June 20 th to June 22 nd 2018,and totally 10 800 pictures were shot during these three days. Five shooting times and six kinds of weed densities were took into consideration for each day. The experimental results showed that the mean identification accuracy was 98. 93%,and neither the shooting times( P > 0. 05) nor the weed densities( P >0. 05) had significant influence on the recognition accuracies. The research result had applicable value,and it can be used as the upstream technology for the intelligent agricultural equipment.
作者
刘慧力
贾洪雷
王刚
GLATZEL Stephan
袁洪方
黄东岩
LIU Huili;JIA Honglei;WANG Gang;GLATZEL Stephan;YUAN Hongfang;HUANG Dongyan(College of Biological and Agricultural Engineering,Jilin University,Changchun 130022,China;Key Laboratory of Bionic Engineering,Ministry of Education,Jilin University,Changchun 130022,China;Department of Geography and Regional Research,University of Vienna,Vienna 1090,Austria)
出处
《农业机械学报》
EI
CAS
CSCD
北大核心
2020年第4期207-215,共9页
Transactions of the Chinese Society for Agricultural Machinery
基金
吉林省科技发展计划国际科技合作项目(20180414074GH)
国家重点研发计划项目(2017YFD0700904)。
关键词
玉米秧苗
茎秆识别
深度学习
膨胀卷积
图像处理
maize seedling
stems recognition
deep learning
expansion convolution
image processing
