基于深度卷积神经网络的水稻病害图像识别研究  

IMAGE RECOGNITION OF RICE DISEASES BASED ON DEEP CONVOLUTIONAL NEURAL NETWORK

作　　者：谭云兰[1,2] 欧阳春娟[1,2] 李龙 廖婷 汤鹏杰 TAN Yun-lan;OUYANG Chun-juan;LI Long;LIAO Ting;TANG Peng-jie(School of Electronic Information and Engineering,Jinggangshan University,Ji’an,Jiangxi 343009,China;Key laboratory of watershed ecology and geographical environment monitoring,National Administration of Surveying,Mapping and Geoinformation,Ji’an,Jiangxi 343009,China;School of Mathematical and Physical Science,Jinggangshan University,Ji’an,Jiangxi 343009,China)

机构地区：[1]井冈山大学电子与信息工程学院,江西吉安343009 [2]井冈山大学流域生态与地理环境监测国家测绘地理信息局重点实验室,江西吉安343009 [3]井冈山大学数理学院,江西吉安343009

出　　处：《井冈山大学学报（自然科学版）》2019年第2期31-38,共8页Journal of Jinggangshan University (Natural Science)

基　　金：国家自然科学基金项目(61462046);流域生态与地理环境监测国家测绘地理信息局重点实验室招标项目(WE2016015);江西省教育厅科学技术研究项目(GJJ160750;GJJ170632;GJJ170643);江西省高校人文社会科学重点研究基地招标项目(JD17082);井冈山大学博士科研启动项目(JZB1807)

摘　　要：水稻病害类型多,采集过来的图像病斑交界特征复杂多变。即便同类别水稻病害在不同的生长时期,发生在叶片、茎秆、穗部等部位呈现的病斑特征也不一样,而且不同类型病害也存在相似病斑,这些都给水稻病害图像的精准识别带来了相当大的困难。采用深度卷积神经网络模型,使用数据集扩增技术,运用fine-tune方法对网络进行调参及构建,将自然场景下采集的常见8类水稻病害图像输入网络模型中进行训练和测试,在有限的图像数量下取得较高的识别精度,其中纹枯病的准确率为93%。不同于其他方法仅聚焦在水稻叶部或稻穗部,本文识别的图像是多株水稻的场景,可为水稻病害远程自动诊断提供关键技术支持。There are many types of rice diseases, and the corresponding disease spot features of the collected images are complex and illegible. The reason is that the characteristics of the lesions occurring in leaves, stems, panicles and other parts are different at different growth stages of the same kind of rice diseases, while there are similar lesions in different types of diseases, which bring great difficulties to the accurate recognition of rice disease images. The purpose of this paper is to recognize the typical rice disease. Based on the deep learning technique, the recognition method for eight rice diseases is put forward. Firstly, the rice disease database of 1260 images is established. Secondly, based on the Caffe deep learning platform, the deep convolution neural network model is constructed through fine-tuning. Finally, eight kinds of common rice diseases images by data augmentation, which are collected in natural scenes, are input into the network model for training and testing. The experimental results show that the proposed method can effectively identify main rice diseases. The recognition accuracy of rice sheath blight is 93% under limited images. Unlike other methods which focus only on rice leaves or panicles, the proposed method in this paper can recognized the multi-plant rice scene image. The proposed method can offer a technical support for remote automatic diagnosis of rice diseases.

关 键 词：图像识别 卷积神经网络 水稻病害 GoogLeNet 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]