基于三维重建的多角度葡萄叶病害识别方法研究

doi:10.13304/j.nykjdb.2021.0507

摘要/Abstract

摘要：

为了解决葡萄在生长过程中因病害侵袭导致品质和产量下降的问题，提出了基于三维重建的多角度图像识别模型。该模型通过三维建模技术对数据进行增强，并扩充数据集用于特征辅助训练，最后与卷积神经网络相结合实现对葡萄叶片病害的识别。在测试集上，训练的3D-MobileNet、3D-Darknet53、3D-resnet34和3D-Resnet101模型相比原模型对葡萄叶片病害识别的准确率分别提高了7.2%、9.6%、10.2%、19.1%。结果表明，提出的基于三维的多角度葡萄叶片病害识别方法能够有效识别葡萄叶病害，为实现葡萄病害的自动识别提供参考。

关键词: 三维重建, 特征辅助训练, 病害识别, 卷积神经网络, 识别准确率

Abstract:

In order to solve the problem of quality and yield decline caused by disease invasion in grape growth， this paper presented a multi-angle image recognition model based on three-dimensional. The model enhanced the data and expanded the dataset for feature aided training by 3D modeling technology. Finally， it combined with convolutional neural network to realize the identification of grape leaf diseases. The accuracy of 3D-MobileNet， 3D-Darknet53， 3D-Resnet34 and 3D-Resnet101 models increased 7.2%， 9.6%， 10.2% and 19.1% than the original model， respectively. The results showed that the method based on 3D multi-angle grape leaf disease identification could effectively identify grape disease types， and provided reference for automatic recognition of grape diseases.

Key words: three-dimensional reconstruction, feature aided training, identify disease, convolutional neural network, recognition accuracy

中图分类号:

S126

方逵, 李成, 何潇, 陈益能. 基于三维重建的多角度葡萄叶病害识别方法研究[J]. 中国农业科技导报, 2022, 24(7): 86-96.

Kui FANG, Cheng LI, Xiao HE, Yineng CHEN. Research on Multi-angle Identification of Grape Leaf Disease Based on 3D Reconstruction[J]. Journal of Agricultural Science and Technology, 2022, 24(7): 86-96.

图/表 14

参考文献 17

1	SHANTKUMARI M, UMA S V. Grape leaf segmentation for disease identification through adaptive Snake algorithm model [J]. Multimed. Tools Appl., 2021, 80(6): 8861-8879.
2	MA J, DU K, ZHENG F, et al.. A recognition method for cucumber diseases using leaf symptom images based on deep convolutional neural network [J]. Comput. Electron. Agric.,2018,154:18-24.
3	王守志,何东健,李文,等.基于核K-均值聚类算法的植物叶部病害识别[J].农业机械学报,2009,40(3):152-155.
	WANG S Z, HE D J, LI W, et al.. Plant leaf disease recognition based on kernel K-means clustering algorithm [J]. Trans. Chin. Soc. Agric. Mach., 2009,40(3):152-155.
4	李颀,赵洁,杨柳,等.基于GA-BP神经网络和特征向量优化组合的黄瓜叶片病斑识别[J].浙江农业学报,2019,31(3):487-495.
	LI Q, ZHAO J, YANG L, et al..Cucumber leaf lesion identification based on GA-BP neural network and feature vector optimization combination [J]. Acta Agric. Zhejiangensis, 2019, 31(3):487-495.
5	ISLAM M, DINH A, WAHID K, et al., Detection of potato diseases using image segmentation and multiclass support vector machine [C]// Proceedings of 2017 IEEE 30th Canadian Conference on Electrical and Computer Engineering (CCECE). IEEE, 2017: 1-4.
6	王翔宇,朱晨光,傅泽田,等.基于可见光光谱分析的黄瓜白粉病识别研究[J].光谱学与光谱分析,2019,39(6):1864-1869.
	WANG X Y, ZHU C G, FU Z T, et al.. Analysis on susceptibility of vegetation canopy spectra in coal mining area to land reclamation [J]. Spectroscopy Soectral Anal., 2019,39(6):1864-1869.
7	ZHU J, WU A, WANG X, et al.. Identification of grape diseases using image analysis and BP neural networks [J]. Multimedia Tools Appl., 2020, 79(21): 14539-14551.
8	LIU B, DING Z, TIAN L, et al.. Grape leaf disease identification using improved deep convolutional neural networks [J/OL]. Front. Plant Sci., 2020, 11: 1082 [2022-05-15]. .
9	苏仕芳, 乔焰, 饶元. 基于迁移学习的葡萄叶片病害识别及移动端应用[J]. 农业工程学报, 2021, 37(10):127-134.
	SU S F, QIAO Y, RAO Y. Recognition of grape leaf diseases and mobile application based on transfer learning [J]. Trans. Chin. Soc. Agric. Eng., 2021, 37(10):127-134.
10	JI M, ZHANG L, WU Q. Automatic grape leaf diseases identification via UnitedModel based on multiple convolutional neural networks [J]. Inform. Proc. Agric., 2020, 7(3): 418-426.
11	刘洋,冯全,王书志.基于轻量级CNN的植物病害识别方法及移动端应用[J].农业工程学报,2019,35(17):194-204.
	LIU Y, FENG Q, WANG S Z. Plant disease identification method based on lightweight CNN and mobile application [J]. Trans. Chin. Soc. Mach., 2019,35(17):194-204.
12	何欣,李书琴,刘斌.基于多尺度残差神经网络的葡萄叶片病害识别[J].计算机工程,2021,47(5):285-291, 300.
	HE X, LI S Q, LIU B. Identification of grape lleaf diseases based on multi-scale residual neural network [J]. Comput. Eng., 2021,47(5):285-291, 300.
13	LIU B, TAN C, LI S, et al.. A data augmentation method based on generative adversarial networks for grape leaf disease identification [J]. IEEE Access, 2020, 8: 102188-102198.
14	赖亦斌. 基于三维点云的小型植物表型提取研究[D].桂林:广西师范大学,2021.
	LAI Y B. 3D point cloud-based small plant phenotypic characteristics extraction [D]. Guilin: Guangxi Nomal University, 2021.
15	HE J Q, HARRISON R J, LI B. A novel 3D imaging system for strawberry phenotyping [J]. Plant Methods, 2017, 13(1): 1-8.
16	MOULON P, MONASSE P, PERROT R, et al.. Openmvg: open multiple view geometry [C]// Proceedings of International Workshop on Reproducible Research in Pattern Recognition. Springer, Cham, 2016: 60-74.
17	CIGNONI P, CALLIERI M, CORSINI M, et al.. Meshlab: an open-source mesh processing tool [C]// Proceedings of Eurographics Italian Chapter Conference. Eurographics, 2008: 129-136.

类别 Category	初始训练集 Initial training set	测试集 Test set	验证集 Validation set	Mag-3D
褐斑病 Brown spot	123	40	40	7 827
气灼病 Gas burning	55	18	18	3 520
健康叶片 Healthy leaf	116	41	41	7 424
缺镁症 Magnesium deficiency	81	27	27	5 184
白粉病 Powdery mildew	139	47	47	8 896

类别 Category	初始训练集 Initial training set	测试集 Test set	验证集 Validation set	Mag-3D
褐斑病 Brown spot	123	40	40	7 827
气灼病 Gas burning	55	18	18	3 520
健康叶片 Healthy leaf	116	41	41	7 424
缺镁症 Magnesium deficiency	81	27	27	5 184
白粉病 Powdery mildew	139	47	47	8 896

模型Model	召回率Recall/%					精确率Precision/%
模型Model	褐斑病 Brown spot	白粉病 Powdery mildew	气灼病 Gas burning	缺镁症 Magnesium deficiency	健康 Health	褐斑病 Brown spot	白粉病 Powdery mildew	气灼病 Gas burning	缺镁症 Magnesium deficiency	健康 Health
MobileNet	27.5	97.9	83.3	71.4	92.7	78.6	63.0	88.2	100.0	79.2
16-MobileNet	32.5	89.4	72.2	71.4	97.6	68.4	64.6	86.7	100.0	75.5
3D-MobileNet	65.0	95.7	88.9	81.0	97.6	92.9	80.4	88.9	94.4	85.1
Darknet53	30.0	89.4	83.3	71.4	100.0	70.6	64.6	88.2	100.0	77.4
16-Darknet53	37.5	85.1	77.8	85.7	97.6	68.2	66.7	93.3	100.0	76.9
3D-Darknet53	70.0	89.4	83.3	81.0	87.8	77.8	80.8	100.0	89.5	80.0
Resnet34	30.0	100.0	66.7	76.2	95.1	85.7	61.0	85.7	100.0	84.8
16-Resbet34	62.5	93.6	66.7	90.5	100.0	86.	80.0	100.0	100.0	78.8
3D-resnet34	75.0	80.9	100.0	85.7	100.0	85.7	88.4	75.0	100.0	87.2
Resnet101	17.5	97.9	61.1	52.4	87.8	87.5	52.9	68.8	100.0	80.0
16-Resnet101	67.5	93.6	72.2	66.7	100.0	87.1	75.9	86.7	93.3	85.4
3D-Resnet101	67.5	91.5	88.9	66.7	97.6	84.4	81.1	80.0	93.3	85.1

模型Model	召回率Recall/%					精确率Precision/%
模型Model	褐斑病 Brown spot	白粉病 Powdery mildew	气灼病 Gas burning	缺镁症 Magnesium deficiency	健康 Health	褐斑病 Brown spot	白粉病 Powdery mildew	气灼病 Gas burning	缺镁症 Magnesium deficiency	健康 Health
MobileNet	27.5	97.9	83.3	71.4	92.7	78.6	63.0	88.2	100.0	79.2
16-MobileNet	32.5	89.4	72.2	71.4	97.6	68.4	64.6	86.7	100.0	75.5
3D-MobileNet	65.0	95.7	88.9	81.0	97.6	92.9	80.4	88.9	94.4	85.1
Darknet53	30.0	89.4	83.3	71.4	100.0	70.6	64.6	88.2	100.0	77.4
16-Darknet53	37.5	85.1	77.8	85.7	97.6	68.2	66.7	93.3	100.0	76.9
3D-Darknet53	70.0	89.4	83.3	81.0	87.8	77.8	80.8	100.0	89.5	80.0
Resnet34	30.0	100.0	66.7	76.2	95.1	85.7	61.0	85.7	100.0	84.8
16-Resbet34	62.5	93.6	66.7	90.5	100.0	86.	80.0	100.0	100.0	78.8
3D-resnet34	75.0	80.9	100.0	85.7	100.0	85.7	88.4	75.0	100.0	87.2
Resnet101	17.5	97.9	61.1	52.4	87.8	87.5	52.9	68.8	100.0	80.0
16-Resnet101	67.5	93.6	72.2	66.7	100.0	87.1	75.9	86.7	93.3	85.4
3D-Resnet101	67.5	91.5	88.9	66.7	97.6	84.4	81.1	80.0	93.3	85.1

模型Model	召回率Recall/%					精确率Precision/%
模型Model	褐斑病 Brown spot	白粉病 Powdery mildew	气灼病 Gas burning	缺镁症 Magnesium deficiency	健康 Health	褐斑病 Brown spot	白粉病 Powdery mildew	气灼病 Gas burning	缺镁症 Magnesium deficiency	健康 Health
MobileNet	45.0	97.9	61.1	76.2	97.6	45.0	97.9	61.1	76.2	97.6
16-MobileNet	70.0	93.6	83.3	85.7	97.6	70.0	93.6	83.3	85.7	97.6
3D-MobileNet	80.0	89.4	94.4	85.7	97.6	80.0	89.4	94.4	85.7	97.6
Darknet53	57.5	95.7	83.3	71.4	95.1	57.5	95.7	83.3	71.4	95.1
16-Darknet53	70.0	87.2	83.3	71.4	90.2	70.0	87.2	83.3	71.4	90.2
3D-Darknet53	62.5	95.7	83.3	81.0	95.1	62.5	95.7	83.3	81.0	95.1
Resnet34	50.0	91.5	72.2	85.7	100.0	50.0	91.5	72.2	85.7	100.0
16-Resbet34	80.0	83.0	88.9	81.0	95.1	80.0	83.0	88.9	81.0	95.1
3D-resnet34	82.5	85.1	94.4	90.5	97.6	82.5	85.1	94.4	90.5	97.6
Resnet101	45.0	100.0	33.3	76.2	90.2	45.0	100.0	33.3	76.2	90.2
16-Resnet101	70.0	93.6	94.4	71.4	90.2	70.0	93.6	94.4	71.4	90.2
3D-Resnet101	72.5	78.7	94.4	85.7	100.0	72.5	78.7	94.4	85.7	100.0