| 一种植物病害图像识别卷积网络架构 |
| |
| 引用本文： | 吴云志,刘翱宇,朱小宁,刘晨曦,范国华,乐毅,张友华.一种植物病害图像识别卷积网络架构[J].安徽农业大学学报,2021,48(1):150-156. |
| |
| 作者姓名： | 吴云志 刘翱宇 朱小宁 刘晨曦 范国华 乐毅 张友华 |
| |
| 作者单位： | 安徽农业大学信息与计算机学院,合肥230036;安徽省北斗精准农业信息工程实验室,合肥230036;安徽农业大学信息与计算机学院,合肥230036;安徽农业大学茶与食品科技学院,合肥230036;湖南大学信息科学与工程学院,长沙410082 |
| |
| 基金项目： | 国家重点研发计划课题（2017YFD0301303）, 安徽省科技重大专项（18030901029）, 安徽省高校自然科学研究项目（KJ2019A0211）, 2018年安徽省省级质量工程项目（2018JXTD114）和安徽省大学生创新创业训练计划项目（S202010364210）共同资助。 |
| |
| 摘 要： | 针对人工提取植物病害图像特征存在效率低、识别率低、成本高等问题,提出一种基于DenseNet网络的现代卷积神经网络架构FI-DenseNet,旨在对多种类的植物病害图像达到高精准的识别准确率.引入Focal损失函数对DenseNet网络进行改进,使得训练模型的注意力集中于难分类的样本.FI-DenseNet网络可以增强特征传递、进行深层训练或有效改善过拟合问题.采用的数据集有87 867张植物病害图像,图像包含同种植物的多种病害,并涉及38种植物病害.对图像进行预处理、数据增强后,将DenseNet169网络、ResNet50网络和MobileNet网络作为参照实验.实验结果表明,FI-DenseNet网络的收敛速度更快且识别准确率最高,测试集识别准确率为98.97％,FI-DenseNet网络的鲁棒性和泛化能力均优于对照网络,可为植物病害智能诊断提供参考.
|
| 关 键 词： | DenseNet 植物病害 卷积神经网络 图像识别 |
A convolutional network for plant disease image recognition (FI-DenseNet) |
| |
| WU Yunzhi,LIU Aoyu,ZHU Xiaoning,LIU Chenxi,FAN Guohu,YUE Yi,ZHANG Youhua.A convolutional network for plant disease image recognition (FI-DenseNet)[J].Journal of Anhui Agricultural University,2021,48(1):150-156. |
| |
| Authors: | WU Yunzhi LIU Aoyu ZHU Xiaoning LIU Chenxi FAN Guohu YUE Yi ZHANG Youhua |
| |
| Institution: | School of Information and Computer Science, Anhui Agricultural University, Hefei 230036; Anhui Provincial Engineering Laboratory for Beidou Precision Agriculture Information, Anhui Agricultural University, Hefei 230036;School of Tea and Food Technology, Anhui Agricultural University, Hefei 230036;College of Information Science and Electronic Engineering, Hunan University, Changsha 410082 |
| |
| Abstract: | Nowadays, artificially extracting plant disease image features has problems such as low efficiency, low recognition rate, and high cost. In this paper, we propose a modern convolutional neural network architecture FI-DenseNet based on DenseNet, which achieves high-precision recognition accuracy for various types of plant disease images. We introduce the Focal loss function to improve the DenseNet so that the attention of the training model is focused on the types of samples that are difficult to classify. The FI-DenseNet can enhance feature transfer, perform deep training, and effectively improve overfitting problems. The dataset used in this paper has 87 867 plant disease images. The image contains multiple plant diseases of the same species and involves 38 plant diseases. After preprocessing the image and enhancing the data, we used DenseNet169, ResNet50, and MobileNet as reference experiments. The experimental results show that the FI-DenseNet has faster convergence speed and the highest recognition accuracy rate. The test set recognition accuracy rate is 98.97%. The robustness and generalization ability of the FI-DenseNet are better than the baselines, which can provide a reference for the intelligent diagnosis of plant diseases. |
| |
| Keywords: | DenseNet plant diseases convolutional neural network image recognition |
| 本文献已被 CNKI 万方数据 等数据库收录！ |
| 点击此处可从《安徽农业大学学报》浏览原始摘要信息 |
| 点击此处可从《安徽农业大学学报》下载免费的PDF全文 |
|
