摘要
玉米叶片病害是影响玉米产量的重要因素之一。目前的识别方法受个人经验和传统图像识别技术的限制,难以达到良好的识别效果。文章以玉米锈病、玉米叶枯病、玉米灰斑病3种典型的病害为例,选取PlantVillage公开数据集的500张图像作为每种病害样本,建立了基于VGG16、Inception V3、ResNet50、ResNet101、DenseNet121的5种深度卷积神经网络的病虫害识别模型,保留原始结构卷积层并设计新的全连接层,再利用迁移学习迁移各个模型ImageNet卷积层权重参数,对比5种模型性能选取最优的网络模型。结果表明,经过重新设计全连接层的DenseNet121性能最优,准确率、召回率、特异率分别为99.73%、99.73%和99.87%,与其他模型相比DenseNet121参数量小、训练时间短,3种病害识别精确率分别为99.6%、100%和99.6%,可精准地识别玉米病害。
One of the important factors of corn leaf disease affecting corn yield.The current recognition methods are limited by personal experience and traditional image recognition technology,so it is difficult to achieve good recognition results.Taking three typical diseases of corn rust,corn leaf blight and corn gray spot as examples,this paper selects 500 images of PlantVillage open data set as samples of each disease,establishes five kinds of deep convolutional neural networks based on VGG16,Inception V3,ResNet50,ResNet101 and DenseNet121 to identify diseases and insect pests,retains the original structure convolution layer and designs a new full connection layer,and then uses transfer learning to transfer the weight parameters of each model ImageNet convolution layer.The best network model is selected by comparing the performance of the five models.The results show that the DenseNet121 performance of the redesigned full connection layer is the best,and the accuracy,recall rate and specificity rate are 99.73%,99.73%and 99.87%respectively.Compared with other models,the number of DenseNet parameters is small,the training time is short,and the accuracy of the three diseases is 99.6%,100%and 99.6%,respectively,thereby can achieve better corn disease recognition.
出处
《智慧农业导刊》
2021年第10期1-10,共10页
JOURNAL OF SMART AGRICULTURE
基金
黑龙江省“百千万”工程科技重大专项(编号:2019ZX14A04)
中央支持地方高校改革发展资金(编号:2020GSP15)
东北农业大学“东农学者计划”项目(编号:19YJXG02)。
作者简介
李恩霖(1997-),男,硕士研究生在读,研究方向:智慧农业、机器视觉;谢秋菊(1976-),女,博士研究生,教授,研究方向:农业信息技术;通信作者:苏中滨(1965-),男,博士研究生,教授,博士研究生导师,研究方向:农业信息技术。
