【作者】 晁晓菲；

【导师】 何东健；

【作者基本信息】 西北农林科技大学 ， 农业电气化与自动化， 2021， 博士

【摘要】 苹果是我国产量最大的水果,而病害是影响我国苹果产业健康发展的重要因素,严重影响苹果的产量和经济效益。苹果叶片常见病害有锈病、灰斑病、斑点落叶病、花叶病、褐斑病等。这些病害引起叶片颜色改变,甚至叶片脱落,叶片大量脱落会削弱果树的树势,降低树体对病害的抵抗能力,导致果实减产或品质下降。因此,快速、准确识别苹果叶片病害及病害程度,对苹果病害精准防治及减少经济损失具有重要的意义。针对基于图像处理的病害识别需要人工挑选图像特征及识别模型普适性和迁移性差等问题。本文以苹果叶片病害图像为研究对象,重点研究基于深度学习的苹果叶片病害识别以及基于深度学习的苹果叶片病害图像语义分割。为实现苹果叶片病害的自动、高效、准确的诊断提供技术支持。论文的主要研究内容及结论如下:(1)提出一种融合Xception及DenseNet网络优点的XDNet深度学习网络。首先,采用角度干扰、光线干扰、噪声干扰以及颜色空间干扰等数据增强技术,对收集的苹果叶片病害数据集进行扩充,建立了包含5种常见叶片病害和健康叶片图像的数据集。然后,结合Xception网络参数少、运算快,而DenseNet网络特征重用性能好的优点,在Xception网络的深层融入DenseNet中的密集连接模块,构建XDNet网络。首先将所有网络进行预训练,然后在收集的苹果叶片病害数据上进行实验。试验结果表明,提出的XDNet网络的病害识别准确率达到98.35%,比Xception、DenseNet,以及其他经典的深度学习分类网络性能更好。且采用的数据增强操作以及迁移学习提升了模型的分型性能和收敛速度。(2)提出SE-DEEP模块,将具有通道注意力的SE(Squeeze and Excitation,SE)模块与深度可分离卷积融合,并将SE-DEEP模块融入Xception网络构建SE_Xception网络,然后通过实验对其进行轻量化,设计了SE_mini Xception网络。针对Xception网络中深度可分离卷积的特点,在SE模块作者所提出的4种SE模块与其他CNN网络融合的方案之外,提出了SE-DEEP模块,将SE模块与深度可分离卷积融合,即将SE模块插入到深度可分离卷积的深度卷积与点卷积之间,并采用SE-DEEP模块对Xception网络进行改进,提出了SE_Xception网络,其病害识别准确率达到99.40%。考虑到病害识别模型在移动端部署的需要,通过压缩模型深度和宽度的方法对SE_Xception网络进行了压缩实验,通过实验对比设计了SE_mini Xception网络,实验证明SE_mini Xception的病害识别准确率达到97.01%,高于Mobile Net与ShuffleNet,模型压缩比达到69:1,满足网络在移动端部署的参数量及运算量的需求。(3)对比了U-Net、DeepLabV3+、PSPNet、SegNet等语义分割网络在苹果叶部病斑分割中的性能,优选了用于苹果叶片病斑分割的网络。采用迁移学习与数据增强策略,在不同超参数组合条件下,对上述4种网络进行试验,以得到较优的超参数组合。试验结果表明,U-Net在苹果叶片病害图像语义分割任务中具有最好的性能,苹果叶部病斑分割的平均交并比达到93.21%。选用ImageNet数据集上预训练的MobileNetV1作为U-Net网络的骨干网络,可取得最好的病害分割结果。(4)融合全局空间注意力的苹果叶片病害图像语义分割网络构建。提出了斜纹池化方法和十字交叉池化方法,构建了提升网络全局空间注意力提取能力的斜纹池化模块(Twill Pooling Module,TPM)、十字交叉池化模块(Double Cross Pooling Module,DCPM)以及十字交叉混合池化模块(Double Cross Mixed Pooling Module,DCMPM),并将所设计的模块融入U-Net,构建了DCPU-Net网络。将各模块加入到U-Net网络进行实验,为不同位置选择合适模块融入所设计的DCPU-Net网络。在与U-Net融合实验中,所提出的TPM、DCPM比SPM性能更好,DCMPM模块也比MPM模块的分割性能更好。设计的DCPU-Net平均交并比达到95.46%,比U-Net提升了2.09%。(5)苹果叶片病害识别与病害程度诊断软件系统框架设计。采用B/S模式,前端采用Vue框架并使用HTML、CSS、Javascript进行编写,后端基于Spring Boot框架,采用Java语言进行开发,利用所设计DCPU-Net为后台模型,实现对病害的识别与病斑语义分割,并在病斑分割结果基础上,依据国家标准进行病害程度诊断。用户可输入指定果园区域抽样点的病害图像,系统可计算一个果园区域的病情指数,并为用户提供防治建议。系统还具有果农交流论坛以及常见病害防治科普等功能。主要功能测试结果表明,该系统界面简洁易用,能够实现预期的病害识别与病害程度诊断功能。病害程度及果园病情估计功能测试结果表明,系统的病害识别准确率达到96.52%,病害程度诊断准确率达到85.81%,果园区域病情估计模块能够实现预定功能且能够对各种特殊情况作出处理。更多还原

【Abstract】 Apple is the fruit with the largest yield in China,and disease is an important factor affecting the healthy development of apple industry in China,which seriously affects the yield and economic benefits of apple industry.The main diseases of apple are Alternaria leaf spot disease,powdery mildew,rust,black star disease,virus disease,silver leaf disease and so on.These diseases could cause leaf color change and even leaf shedding.A large number of leaf shedding will weaken the tree strength of the apple tree,reduce the resistance of the tree to diseases,and lead to the decline of fruit yield or quality.Therefore,the rapid and accurate identification of apple leaf diseases and disease degree is of great significance for the precise prevention and control of apple disease and the reduction of economic losses.In order to overcome the shortage of disease recognition based on image processing methods,which manually select image features,and their recognition model has poor transferability.In this paper,apple leaf disease images are taken as the research object,and the identification of apple leaf disease and semantic segmentation of apple leaf images based on deep learning are mainly studied.It provides technical support for automatic,efficient and accurate diagnosis of apple leaf diseases.The major research contents and conclusions of this paper are as follows:(1)XDNet deep learning network combining the advantages of Xception and DenseNet networks is proposed.Firstly,a dataset containing images of 5 common leaf diseases and healthy leaves was established,and the collected apple leaf disease dataset was expanded by using angle interference,light interference,noise interference and color space interference.Then,considering that Xception has fewer parameters and fast operation speed because of the usage of depth-wise separable convolutions,and the Dense block in DenseNet enable the network more powerful in feature reusing ability,we proposed XDNet network by incorporating dense connection modules into the deep layer of Xception network,Combined with the advantages of Xception network and DenseNet.All the networks were pre-trained on subset of Plant Village dataset,and then transferred to the collected apple leaf diseases dataset.Experimental results show that the proposed XDNet network has a disease recognition accuracy of 98.35%,which is better than Xception,Densenet,and other classic deep learning classification networks.The data augmentation and transfer learning technologies improved the classification performance and convergence speed of the model.(2)In order to introduce channel attentions into convolutional networks,SE-DEEP module to fuse SE(Squeeze and Excitation)module with depth-wise separable convolution was proposed.Then,SE-DEEP module was combined with Xception network to put forward the SE_Xception network.Then,SE_mini Xception network was designed to compress SE_Xception network through experimets.Since depth-wise separable convolution can be separated into the depth-wise convolution and pointwise convolution,this paper proposed a method to insert SE module into Xception network,which is different from those four methods for SE module integrated into CNN networks proposed in the paper which the SE module was proposed.In the proposed method,SE module is insert between the depth-wise convolution and pointwise convolution,then based on the deep fusion of SE module with Xception network,SE_Xception is proposed,and its disease recognition accuracy reaches 99.40%.Considering the need of disease recognition model deployment in mobile terminal,the model compression experiment was carried out by squeezing the depth and width of the network.By comparing different compression methods and different compression scales,SE_minixception with small number of parameters and better disease recognition performance is designed,its disease recognition accuracy reaches 97.01%,which is higher than Mobile Net and ShuffleNet.The network compression ratio of SE_minixception reaches 69:1,and it meets the requirements of the number of parameters and the amount of computation for network deployment in the mobile terminal.(3)Four sematic segmentation networks were compared for the sematic segmentation of apple leaf disease semantic segmentation task,including U-Net,Deeplabv3 +,PSPNet and Segnet,and the segmentation network for the apple leaf lesion segmentation was selected.Using transfer learning and data augmentation technology,the above four networks are tested under different hyperparameter combinations,and the best hyperparameter combinations is selected for each network.The experimental results show that U-Net has the best average performance in semantic segmentation of apple leaf images,its mean intersection over union reaches 93.21%,and Mobile Net V1 pretrained in Image Net dataset is the best backbone for UNet.(4)In order to add the global spatial attentions to the semantic segmentation network,the apple lesion segmentation network with global spatial context was built.TPM(Twill Pooling Module),DCPM(Double Cross Pooling Module)and DCMPM(Double Cross Mixed Pooling Module)modules were constructed to enable segmentation backbones to efficiently model long-range dependencies.The designed modules were integrated into U-Net,so that the DCPU-Net semantic segmentation network was designed.Based on the analysis of the advantages and possibilities to improve SPM(Strip Pooling Moudle)and MPM(Mixed Pooling Module)modules,the twill pooling module was constructed,and the DCPM module was proposed by combining TPM with SPM.On the basis of MPM,the twill pooling was integrated to get the DCMPM module.Each module was added to the U-Net network for experiment,and appropriate modules were selected for different locations of the designed DCPU-Net network.In the fusion experiment of U-Net with Mobile NetV1 as backbone network,the proposed TPM and DCPM modules have better performance than SPM,and the DCMPM module also performs better than MPM.The mean intersection over union of the designed DCPU-Net reached 95.46%,which was 2.09% higher than that of baseline U-Net.(5)The software framework for apple leaf disease identification and disease degree diagnosis was designed and developed.Uses B/S mode,Vue framework and HTML,CSS,Java script were used for the front-end design,Spring Boot framework and Java language were used for back-end design,and the proposed DCPU-Net was embedded to realize the apple disease identification and leaf lesion segmentation,and then diagnose the disease degree based on the segmentation result according to the national standard for disease degree diagnosis.The user can input the disease images of the sampling points in the designated orchard area,and the system can calculate the disease index of an orchard area,and provide the user with prevention and treatment suggestions.The system also has the functions of communication forum for fruit growers and information browsing of common diseases prevention and control.The main function tests show that the interface of the system is simple and easy to use,and the expected functions of disease identification,disease degree diagnosis and diseae index of orchar area can be realized.Test results show that system disease recognition accuracy is96.52%,and system disease degree diagnosis accuracy is 85.81%,the orchard area disease estimation module can realize the predicted function and can deal with various special cases.更多还原