摘要
为了快速检测芒果树上的芒果,提出了一种基于无人机的树上绿色芒果视觉检测方法。采用深度学习技术,利用YOLOv2模型对无人机采集的绿色芒果图像进行检测,首先通过无人机采集树上绿色芒果图像,对芒果图像进行人工标记,建立芒果图像的训练集和测试集,通过试验确定训练模型的批处理量和初始学习率,并在训练模型时根据训练次数逐渐降低学习率,最终训练的模型在训练集的平均精度(Mean average precision,MAP)为86. 43%。试验分析了包含不同果实数和不同光照条件下绿色芒果图像的识别正确率,并进行了芒果产量估计试验,试验结果表明:本文算法检测一幅图像的平均运行时间为0. 08 s,对测试集的识别正确率为90. 64%,识别错误率为9. 36%;对含不同果实数的图像识别正确率为88. 05%~94. 55%,顺光条件下识别正确率为93. 42%,逆光条件下识别正确率为87. 18%;对芒果产量估计的平均误差为12. 79%。本文算法对自然环境下树上绿色芒果有较好的检测效果,可为农业智能化生产中果蔬产量的估计提供技术支持。
In order to detect the mango yield on trees rapidly,a green mango visual detection method based on unmanned aerial vehicle(UAV)was proposed.The deep learning technology and the YOLOv2 model were adopted to detect the mango images captured by UAV.Firstly,totally 471 images of the mango on trees were collected by the UAV.To meet the demand of diversity,totally 360 images included different shooting distances and different lighting situations were selected.Among which,300 images were selected randomly as the training set,the other 60 images were used as the test set.Also,the shooting plan of the whole tree was designed.By image collecting and image mosaic,the integrated images of five mango trees were worked out for the yield estimating experiment of mango.After image collection,these images were marked manually and used to build the training set and the test set.The batch size and the initial learning rate were determined by experiments.During the model training,the learning rate was reduced gradually as the training times were changed.The mean average precision(MAP)of the trained model on the training set was 86.43%.By designing the experiments,the accuracy of mango recognition with images that containing different fruit numbers and different lighting conditions was worked out.Also,the yield estimation experiment was designed.The experimental results showed that the average running time of an image using the given algorithm was 0.08 s,while the accuracy of the teat set was 90.64%and the false recognition rate was 9.36%;the highest recognition accuracy of image with different numbers of fruits was 94.55%and the lowest was 88.05%.The recognition accuracy was 93.42%under the condition of direct sunlight,and the recognition accuracy was 87.18%under the condition of backlight.The average error of the yield estimation of mango tree was 12.79%.The result demonstrated that the algorithm was effective for mango in natural environment,which can provide technical support for estimating the yield of fruits and vegetables in intelligent agricultural production.
作者
熊俊涛
刘振
林睿
陈淑绵
陈伟杰
杨振刚
XIONG Juntao;LIU Zhen;LIN Rui;CHEN Shumian;CHEN Weijie;YANG Zhengang(College of Mathematics and Informatics,South China Agricultural University,Guangzhou 510642,China)
出处
《农业机械学报》
EI
CAS
CSCD
北大核心
2018年第11期23-29,共7页
Transactions of the Chinese Society for Agricultural Machinery
基金
国家自然科学基金项目(31201135
31571568)
广东省自然科学基金项目(2018A030313330)
广州市科技计划项目(201802020032)
关键词
无人机
绿色芒果
深度学习技术
视觉检测
unmanned aerial vehicle
green mango
deep learning technique
visual detection