一种基于改进ResNet18神经网络的苹果叶片病害识别方法  

A Novel Apple Leaf Disease Recognition Method Based on Improved ResNet18 Neural Network

作　　者：陈诗瑶 孔淳 冯峰 孙博 王志军[1,2] Chen Shiyao;Kong Chun;Feng Feng;Sun Bo;Wang Zhijun(College of Information Science and Engineering,Shandong Agricultural University,Taian 271018,China;Apple Technology Innovation Center of Shandong Province,Taian 271018,China)

机构地区：[1]山东农业大学信息科学与工程学院,山东泰安271018 [2]山东省苹果技术创新中心,山东泰安271018

出　　处：《山东农业科学》2024年第10期174-180,共7页Shandong Agricultural Sciences

基　　金：山东省重大科技创新工程项目“现代果园智慧种植装备与大数据平台研发及示范应用”(2019JZZY010706)。

摘　　要：为有效提升苹果叶片病害识别的精度和效率,实现病害的及时防治进而提高苹果产量,本研究提出一种基于改进ResNet18神经网络的苹果叶片病害识别方法,可在提升模型识别性能的同时减少参数量和模型尺寸。首先,改进ResNet模型的残差结构,以减少参数量,实现模型轻量化;其次,引入坐标注意力(CA)机制并进行迁移学习,进一步提升模型的泛化性能。将改进ResNet18模型与原始ResNet18神经网络进行对比实验,结果发现,改进模型的准确率提升了1.53个百分点,但模型参数量减少为原始模型的50.84%。表明本研究提出的改进ResNet18模型可有效识别苹果叶片病害,且方便移动端搭载。In order to effectively improve the accuracy and efficiency of apple leaf disease recognition,and achieve timely prevention and treatment of apple diseases so as to improve yield,this study proposed an apple leaf disease recognition method based on the improved ResNet18 neural network,which could improved the recognition performance of the model but reduce the parameter quantity and model size.First,improved the residual structure of ResNet network to reduce the parameter quantity,which could achieve the model lightweighting.Second,integrated the coordinate attention(CA)mechanism and transfer learning into the model to further improve its generalization performance.Comparing with the original ResNet18 model,the accuracy of the improved model increased by 1.53 percentage points,but the parameter quantity reduced to 50.84%of the original model.The above results indicated that the improved model could effectively recognize apple leaf diseases and was easy to carry on mobile terminal.

关 键 词：苹果叶片病害识别 卷积神经网络 ResNet18模型 残差结构 坐标注意力机制 迁移学习 

分 类 号：S661.1[农业科学—果树学] S126[农业科学—园艺学]