| 基于深度学习的大豆生长期叶片缺素症状检测方法 |
| |
| 引用本文： | 熊俊涛,戴森鑫,区炯洪,林筱芸,黄琼海,杨振刚.基于深度学习的大豆生长期叶片缺素症状检测方法[J].农业机械学报,2020,51(1):195-202. |
| |
| 作者姓名： | 熊俊涛 戴森鑫 区炯洪 林筱芸 黄琼海 杨振刚 |
| |
| 作者单位： | 华南农业大学数学与信息学院,广州510642;华南农业大学电子工程学院,广州510642 |
| |
| 基金项目： | 广东省重点研发计划项目（2019B020223002）、广东省自然科学基金项目（2018A030313330）、广东省省级大学生创新创业训练计划项目（201810564098）和广东省大学生科技创新培育专项资金项目（Pdjh2018b0079） |
| |
| 摘 要： | 为了检测作物叶片缺素,提出了一种基于神经网络的大豆叶片缺素视觉检测方法。在对大豆缺素叶片进行特征分析后,采用深度学习技术,利用Mask R-CNN模型对固定摄像头采集的叶片图像进行分割,以去除背景特征,并利用VGG16模型进行缺素分类。首先通过摄像头采集水培大豆叶片图像,对大豆叶片图像进行人工标记,建立大豆叶片图像分割任务的训练集和测试集,通过预训练确定模型的初始参数,并使用较低的学习率训练Mask RCNN模型,训练后的模型在测试集上对背景遮挡的大豆单叶片和多叶片分割的马修斯相关系数分别达到了0.847和0.788。通过预训练确定模型的初始参数,使用训练全连接层的方法训练VGG16模型,训练的模型在测试集上的分类准确率为89.42%。通过将特征明显的叶片归类为两类缺氮特征和4类缺磷特征,分析讨论了模型的不足之处。本文算法检测一幅100万像素的图像平均运行时间为0.8 s,且对复杂背景下大豆叶片缺素分类有较好的检测效果,可为农业自动化生产中植株缺素情况估计提供技术支持。
|
| 关 键 词： | 大豆叶片 缺素 深度学习 神经网络 迁移学习 |
| 收稿时间： | 2019/5/4 0:00:00 |
Leaf Deficiency Symptoms Detection Method of Soybean Based on Deep Learning |
| |
| XIONG Juntao,DAI Senxin,OU Jionghong,LIN Xiaoyun,HUANG Qionghai,YANG Zhen gang.Leaf Deficiency Symptoms Detection Method of Soybean Based on Deep Learning[J].Transactions of the Chinese Society of Agricultural Machinery,2020,51(1):195-202. |
| |
| Authors: | XIONG Juntao DAI Senxin OU Jionghong LIN Xiaoyun HUANG Qionghai YANG Zhen gang |
| |
| Institution: | South China Agricultural University,South China Agricultural University,South China Agricultural University,South China Agricultural University,South China Agricultural University and South China Agricultural University |
| |
| Abstract: | In order to detect plant leaf element deficiency,a visual detection method of soybean leaf element deficiency based on neural network was proposed.After analyzing the characteristics of soybean deficient leaves,deep learning technology was used.The Mask R-CNN model was used to segment the leaf images collected by a fixed camera,and the VGG16 model was adopted to classify the deficient leaves.Firstly,after collecting hydroponic soybean images,the outline of soybean leaves was marked manually in the images,establishing training set and test set of segmentation task.Through pre-training,the initial parameters of the Mask R-CNN model were determined,and then using lower learning rate to train the model.For segmentation task on single leaf images and on multiple leaves on a complex background in the test set,the Matthews correlation coefficient(MCC)of the final trained model reached0.847 and 0.788 respectively.The training set and test set of the soybean leaf image classification task were established by segmenting the leaves through the trained Mask R-CNN network and manually marking them.The initial parameters of the VGG16 model were determined through pre-training,and then the whole connection layers of the VGG16 model were replaced before training to adapt to the leaf classification task.The classification accuracy of the final trained model on the test set was 89.42%.When analyzing the result,the leaves with obvious deficiency features were classified into two types of nitrogen deficiency and four types of phosphorus deficiency to discuss the inadequacy of the method.The average running time of the algorithm to detect a picture of 1 million pixels was 0.8 s.The algorithm had a good detection result on the classification of soybean leaf deficiency under complex background,which can provide technical support for the estimation of plant deficiency in agricultural automation production. |
| |
| Keywords: | soybean leaf element deficiency deep learning nerual network transfer learning |
| 本文献已被 CNKI 维普 万方数据 等数据库收录！ |
| 点击此处可从《农业机械学报》浏览原始摘要信息 |
| 点击此处可从《农业机械学报》下载免费的PDF全文 |
|
