[1]龚 安,井晓萌.多卷积神经网络模型融合的农作物病害图像识别[J].计算机技术与发展,2020,30(08):134-139.[doi:10. 3969 / j. issn. 1673-629X. 2020. 08. 023]
　GONG An,JING Xiao-meng.Image Recognition of Crop Diseases Based on Multi-convolution Neural Network Model Ensemble[J].,2020,30(08):134-139.[doi:10. 3969 / j. issn. 1673-629X. 2020. 08. 023]
点击复制

多卷积神经网络模型融合的农作物病害图像识别()
分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
30
期数:
2020年08期
页码:
134-139
栏目:
应用开发研究
出版日期:
2020-08-10

文章信息/Info

Title:
Image Recognition of Crop Diseases Based on Multi-convolution Neural Network Model Ensemble
文章编号:
1673-629X(2020)08-0134-06
作者:
龚 安; 井晓萌
中国石油大学(华东) 计算机科学与技术学院,山东 青岛 266580
Author(s):
GONG An; JING Xiao-meng
School of Computer and Communication Technology,China University of Petroleum (East China),Qingdao 266580,China
关键词:
农作物病害识别; 模型融合; 卷积神经网络; 元学习器; 迁移学习
Keywords:
crop disease identification; model ensemble; convolutional neural network; meta learner; transfer learning
分类号:
TP391
DOI:
10. 3969 / j. issn. 1673-629X. 2020. 08. 023
摘要:
农作物病害是粮食安全的主要威胁,病害的诊断对于农业生产来说至关重要。 针对单一卷积神经网络在农作物病害识别上的局限性,分类准确率不高的问题,采用多个卷积神经网络模型融合的方式,对 10 种农作物的 27 种病害及其3 种病害程度的农作物叶子图片进行病害及病害程度的细粒度识别。首先选用 Resnet101、RestNext50、SE-ResNet50、 SERestNext50 这4 种网络模型运用迁移学习的方式,固定底层模型参数,修改顶层的全连接层进行训练,然后采用 Stacking方法将模型预测结果输入第二层元学习器 XGBoost, 最后对单模型预测结果和 Stacking 融合后的结果进行对比。 实验结果表明,经过模型融合后的准确率能达到 87. 19% ,具有较高的识别准确率及较强的鲁棒性,可以作为农作物病害的早期诊断方式,并可以进一步研究将该方法应用到真实的农业生产中。
Abstract:
Crop disease is a major threat to food security and its diagnosis is of significant importance for agricultural production. For the limitation of a single convolutional neural network in the identification of crop diseases and the problem of low classification accuracy,we adopt the ensemble method of multiple convolution neural network models to identify 27 diseases of 10 crops and their 3 degrees of disease by fine particle size. Firstly,the four network models Resnet101,RestNext50,SE-ResNet50 and SE-RestNext50 are used to transfer learning,fix the underlying model parameters,and modify the top-level fully connected layer for training. Then, the model prediction result is input into the second layer element learner XGBoost by Stacking,and finally the single model prediction result and the stacking result are compared. The experiments show that the accuracy of model fusion can reach 87.19% ,which has high recognition accuracy and strong robustness. It can be used as an early diagnosis method of crop dise-ases,and can be further studied and applied to real agricultural production.

相似文献/References:

[1]余 威,龙慧云.基于深度卷积网络的遥感影像建筑物分割方法[J].计算机技术与发展,2019,29(06):57.[doi:10. 3969 / j. issn. 1673-629X. 2019. 06. 012]
　YU Wei,LONG Hui-yun.A Building Segmentation Method Based on Deep Convolution Networks for Remote Sensing Imagery[J].,2019,29(08):57.[doi:10. 3969 / j. issn. 1673-629X. 2019. 06. 012]
[2]李春生,曹 琦,于 澍.针对小规模数据集的多模型融合算法研究[J].计算机技术与发展,2020,30(02):63.[doi:10. 3969 / j. issn. 1673-629X. 2020. 02. 013]
　LI Chun-sheng,CAO Qi,YU Shu.Research on Multi-model Fusion Algorithm for Small Scale Data Sets[J].,2020,30(08):63.[doi:10. 3969 / j. issn. 1673-629X. 2020. 02. 013]
[3]林益文,杨 啸,陈 青,等.基于距离损失函数的特征融合模型[J].计算机技术与发展,2023,33(12):72.[doi:10. 3969 / j. issn. 1673-629X. 2023. 12. 010]
　LIN Yi-wen,YANG Xiao,CHEN Qing,et al.Feature Fusion Model Based on Distance Loss Function[J].,2023,33(08):72.[doi:10. 3969 / j. issn. 1673-629X. 2023. 12. 010]

更新日期/Last Update: 2020-08-10