基于对比学习的植物叶片病害识别  

Plant leaf disease identification based on contrastive learning

作　　者：杨新宇 冯全 张建华[2] 杨森 YANG Xinyu;FENG Quan;ZHANG Jianhua;YANG Sen(Mechanical and Electrical Engineering College,Gansu Agricultural University,Lanzhou 730070,China;Agricultural Information Institute,Chinese Academy of Agricultural Sciences,Beijing 100081,China)

机构地区：[1]甘肃农业大学机电工程学院,甘肃兰州730070 [2]中国农业科学院农业信息研究所,北京100081

出　　处：《浙江农业学报》2024年第1期215-224,共10页Acta Agriculturae Zhejiangensis

基　　金：国家自然科学基金(32160421,31971792,32201663);甘肃省高等学校产业支撑计划(2021CYZC-57)。

摘　　要：目前,基于图像处理的植物病害识别多依赖于人工标注的卷积神经网络。本文基于自监督对比学习不依赖标签和大量数据即可实现独自学习的优势,研究了MoCo-v2、DeepCluster-v2、SwAV、BYOL这4种对比学习方法对植物叶片病害的识别效果,通过设置不同试验条件,使用PlantVillage开源数据集和自建的棉花病害数据集分别测试4种对比学习方法所训练的ResNet50编码器在Linear和Finetune两种模式下的病害识别效果,评估对比学习方法在植物叶片病害识别上的可行性。结果表明,在PlantVillage数据集中,Finetune模式的平均准确率略高于Linear,4种方法训练的编码器的平均识别准确率最高达99.83%,其中,DeepCluster-v2、BYOL在Finetune模式下的识别率最高,均为99.87%。在棉花病害数据集上,Finetune模式的效果略逊于Linear,DeepCluster-v2在Linear模式下获得最高识别准确率(98.86%)。整体来看,基于对比学习方法的病害识别率优于有监督模型的学习效果,在植物叶片病害识别领域展现出良好的应用前景。At present,the recognition of plant disease via image processing mostly relies on the manually labeled convolutional neural network.However,the self-monitoring contrastive learning could achieve independent learning without relying on labels and large amounts of data.In view of this advantage,the effect of four contrastive learning methods,MoCo-v2,DeepCluster-v2,SwAV and BYOL,on the identification of plant leaf diseases were compared by setting different experimental conditions on the open-source dataset of PlantVillage and the self-built cotton disease dataset.The ResNet50 encoder trained by four contrastive learning methods was tested for disease identification both under Linear and Finetune modes,and the feasibility of the contrastive learning methods in identifying plant leaf diseases was evaluated.It was shown that the average accuracy under Finetune mode on the PlantVillage dataset was higher than that under Linear mode,and the highest identificaiton accuracy of the encoders trained by the four methods reached 99.83%.DeepCluster-v2 and BYOL had the highest identification rate under Finetune mode,both of which were 99.87%.On the self-built cotton disease dataset,the performance under Finetune mode was poorer than that under Linear mode,and the highest idenficaiton accuracy of DeepCluster-v2 under Linear mode was 98.86%.Overall,the disease identification rate based on contrastive learning method was superior to the supervised models,demonstrating good application prospects in plant leaf disease identification.

关 键 词：对比学习 病害识别 图像处理 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]