基于改进轻量级深度卷积神经网络的果树叶片分类及病害识别模型设计  

Design of Classification of Fruit Leaves and Disease Identification Model Based on Improved MobileNet-V2

作　　者：买买提·沙吾提[1,2,3] 李荣鹏 蔡和兵 赵明 梁嘉曦 SAWUT Mamat;LI Rongpeng;CAI Hebing;ZHAO Ming;LIANG Jiaxi(College of Geography and Remote Sensing Sciences,Xinjiang University,Urumqi 830017,China;Xinjiang Key Laboratory of Oasis Ecology,Xinjiang University,Urumqi 830017,China;Key Laboratory of Smart City and Environment Modelling of Higher Education Institute,Xinjiang University,Urumqi 830017,China)

机构地区：[1]新疆大学地理与遥感科学学院,乌鲁木齐830017 [2]新疆绿洲生态重点试验室,乌鲁木齐830017 [3]智慧城市与环境建模自治区普通高校重点试验室,乌鲁木齐830017

出　　处：《森林工程》2025年第2期277-287,共11页Forest Engineering

基　　金：新疆自然科学计划项目(2021D01C055);新疆大学国家级大学生创新训练计划项目(202310755002)。

摘　　要：新疆是中国重要的林果产业基地,特色林果业是区域经济发展的重要组成部分。为预防果树病害制约林果业发展,设计一款归一化注意力(normalization-based attention module,NAM)轻量级深度卷积神经网络(MobileNet-V2)果树叶片分类及病害识别模型。其中融入轻量型的归一化注意力机制,提高模型对特征信息的敏感度,使模型关注显著性特征。同时,将L1正则化(L1 regularization或losso)添加到损失函数中,对权重进行稀疏性惩罚,抑制非显著性权重。试验结果表明,在叶片分类中,模型对自构建植物叶片病害识别数据集(Plant Village)、混合数据集的分类结果均表现良好,准确率分别达到97.05%、98.73%、94.91%,具有较好的泛化能力。在病害识别中,MobileNet-V2 NAM模型实现94.55%的识别准确率,高于深度卷积神经网络(AlexNet)、视觉几何群网络(VGG16)经典卷积神经网络(Convolutional Neural Networks,CNN)模型,且模型参数量只有3.56 M。MobileNet-V2 NAM在具有良好准确率同时保持了较低的模型参数量,为深度学习模型嵌入到移动设备提供技术支持。Xinjiang is an important forest and fruit industry base in China,and the characteristic forest and fruit industry is an important part of regional economic development.In order to prevent fruit tree diseases from restricting the development of forest and fruit industry,a MobileNet-V2 NAM fruit tree leaf classification and disease identification model was designed in this study.It incorporated a lightweight normalization-based attention module to improve the model′s sensitivity to feature information and make the model focus on salient features.At the same time,L1 regularization was added to the loss function to penalize the sparsity of the weights and suppress the non-significant weights.The experimental results showed that:in leaf classification,the model performed well in the classification results of self-built,Plant Village,and mixed datasets,with the accuracy rates reaching 97.05%,98.73%,and 94.91%,respectively,and had good generalization ability.In disease identification,the MobileNet-V2 NAM model achieved a recognition accuracy of 94.55%,which was higher than the AlexNet,VGG16 classic CNN models,and the number of parameters of the model was only 3.56M.MobileNet-V2 NAM has good accuracy while maintaining a low amount of model parameters,provides technical support for embedding deep learning models into mobile devices.

关 键 词：新疆 果树分类 病害识别 归一化注意力轻量级深度卷积神经网络(MobileNet-V2 NAM) 归一化注意力机制 

分 类 号：S436.611[农业科学—农业昆虫与害虫防治]