基于改进ResNet18的苹果叶部病害多分类算法研究  

Identification of Apple Leaf Diseases Based on Improved ResNet18

作　　者：姜红花[1] 杨祥海 丁睿柔 王东伟[2] 毛文华[3] 乔永亮 JIANG Honghua;YANG Xianghai;DING Ruirou;WANG Dongwei;MAO Wenhua;QIAO Yongliang(College of Information Science and Engineering,Shandong Agricultural University,Taian 271018,China;College of Mechanical and Electrical Engineering,Qingdao Agricultural University,Qingdao 266109,China;Chinese Academy of Agricultural Mechanization Sciences Group Co.,Ltd.,Beijing 100083,China;Faculty of Engineering,The University of Sydney,Sydney 2006,Australia)

机构地区：[1]山东农业大学信息科学与工程学院,泰安271018 [2]青岛农业大学机电工程学院,青岛266109 [3]中国农业机械化科学研究院集团有限公司,北京100083 [4]悉尼大学工学院,悉尼2006

出　　处：《农业机械学报》2023年第4期295-303,共9页Transactions of the Chinese Society for Agricultural Machinery

基　　金：山东省农业重大应用技术创新项目(SD2019NJ001);山东省重大科技创新工程项目(2019JZZY010716)。

摘　　要：针对传统苹果叶部病害分类方法精准性差、效率低等问题,提出了一种基于改进ResNet18的苹果叶部病害多分类算法。通过在原始ResNet18网络的基础上增加通道与空间注意力机制分支,强化网络对叶部病害区域的特征提取能力,提高病害的识别精度和实时性。为更好地引导网络学习到零散分布的病害斑点的特征,引入特征图随机裁剪分支,不仅实现有限样本空间的扩充,还进一步优化网络结构,提高训练速度。试验以苹果5类常见的叶部病害(黑星病、黑腐病、雪松锈病、灰斑病、白粉病)为主要研究对象,并与主流分类算法模型进行对比。试验结果表明,所提ResNet18 CBAM RC1模型病害分类准确率可达98.25%,高于ResNet18(93.19%)和VGG16(96.13%),能够有效提取叶片病害特征,增强对多类病害的识别,提高识别准确率。此外,模型内存占用量仅为37.44 MB,单幅图像推理时间为9.11 ms,可满足嵌入式设备上果园病害识别的实时性要求。Aiming at the problem that the traditional apple leaf disease classification method has poor accuracy and low efficiency,which affects prevention and cure effect,an improved ResNet18 algorithm was proposed.By adding the branch of channel and spatial attention mechanism to the original ResNet18,the feature extraction ability of the network for leaf disease regions was strengthened to improve the disease recognition accuracy and real⁃time performance.In addition,to better guide the network to learn the features of sporadically distributed disease spots,the feature map random cropping branch was introduced,which not only achieved the expansion of the limited sample space,but also further optimized the network structure and improved the training speed.The experiment was conducted with five common types of apple foliar diseases(black star,black rot,cedar rust,gray spot,and powdery mildew)as the main research objects and compared with the mainstream classification algorithm models for analysis.The experimental results showed that the disease classification accuracy of the proposed ResNet18 CBAM RC1 model can reach 98.25%,which was higher than that of ResNet18(93.19%)and VGG16(96.13%),and can effectively extract leaf disease features,enhance the recognition of multiple types of diseases,and improve the real⁃time recognition capability and accuracy.In addition,the model size was only 37.44 MB and the inference time of a single image was 9.11 ms,which can meet the real⁃time requirements of orchard disease recognition on embedded devices and provide information support for disease prevention and control in digital orchards.

关 键 词：苹果叶部病害 ResNet18 注意力机制 随机裁剪 深度学习 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]