摘要
作物病害识别对于保障农作物的健康生长和农业生产的稳定发展具有至关重要的意义。近年来,诸多研究表明,数据增强技术的引入可显著提升作物病害识别模型的准确率。基于此,采用数据增强技术提高烟草靶斑病病害识别模型的性能。通过图像翻转、灰度调整、亮度调整、色度调整等基础方法,并进一步引入更为先进的MixUp和CutMix数据增强方法,对烟草靶斑病图像数据进行扩充及多样化处理,利用3种主流图像识别模型AlexNet、GoogleNet和ResNet101,系统评估数据增强对烟草靶斑病识别任务的有效性。结果表明,相较于未使用数据增强的对照组,增强后的模型在训练集与测试集上的准确率最高分别提升4.98%和2.21%,训练集和测试集的损失值分别降低10.84%和4.73%。结果表明,数据增强技术的使用可以提升烟草靶斑病病害识别模型的性能,为烟草病害图像识别模型提供了可靠的数据处理方法,也为图像识别模型的应用提供了科学依据。
Crop disease identification is of great significance for ensuring the healthy growth of crops and the stable development of agricultural production.In recent years,many studies have shown that the introduction of data augmentation techniques has significantly improved the accuracy of crop disease recognition models.This study proposes the application of data augmentation techniques to enhance the performance of tobacco target spot disease recognition models.The research employs various data augmentation methods,including image flipping,grayscale adjustment,brightness adjustment and chroma adjustment,as well as MixUp and CutMix data augmentation methods,to expand and diversify the image data of tobacco target spot disease.The data augmentation effects on the tobacco target spot disease image recognition models were verified using mainstream image recognition models,namely AlexNet,GoogleNet,and ResNet101.The results show that after the application of data augmentation,the training set accuracy and test set accuracy of the image recognition models were increased by up to 2.80%and 3.78%,respectively,compared to those without data augmentation.Meanwhile,the training set loss and test set loss were reduced by 10.84%and 4.73%,respectively.The study concludes that the use of data augmentation techniques can improve the performance of tobacco target spot disease recognition models.This method provides a data processing approach for the research of tobacco disease image recognition models and offers a scientific basis for the application of image recognition models.
作者
陈海涛
孙佳照
罗锦舟
丁伟
CHEN Haitao;SUN Jiazhao;LUO Jinzhou;DING Wei(Chongqing Tobacco Branch,China National Tobacco Corporation,Chongqing 400000,China;School of Plant Protection,Southwest University,Chongqing 400715,China)
出处
《植物医学》
2025年第6期77-84,共8页
Plant Health and Medicine
基金
重庆市烟草病虫害监测与防控数智化平台的构建及应用(B20241NY1301)。
关键词
烟草靶斑病
图像识别模型
数据增强
模型性能
tobacco target spot disease
image recognition model
data augmentation
model performance
