基于深度学习的病虫害智能化识别系统  

Intelligent identification system of disease and insect pests based on deep learning

作　　者：陈天娇[1] 曾娟[2] 谢成军[1] 王儒敬[1] 刘万才[2] 张洁[1] 李瑞[1] 陈红波[1] 胡海瀛[1] 董伟 Chen Tianjiao;Zeng Juan;Xie Chengjun;Wang Rujing;Liu Wancai;Zhang Jie;Li Rui;Chen Hongbo;Hu Haiying;Dong Wei(Institute of Intelligent Machines/Hefei Instit utes of Physical Science,Chinese Academy of Sciences,Hefei Anhui 230031,China;National Agro-Technical Extension and Service Center,Beijing 100125,China;Agricultural Economy and In formation Research Institute,Anhui Academy of Agricultural Sciences,Hefei Anhui 230001,China)

机构地区：[1]中国科学院合肥物质科学研究院/合肥智能机械研究所,安徽合肥230031 [2]全国农业技术推广服务中心,北京100125 [3]安徽省农业科学院农业经济与信息研究所,安徽合肥230001

出　　处：《中国植保导刊》2019年第4期26-34,共9页China Plant Protection

基　　金：国家重点研发计划(2018YFD0200300);国家自然基金项目(31671586;61773360;31401293)

摘　　要：我国农作物种植覆盖面广、分散度高,病虫害发生种类多、区域性发生规律复杂,传统的人工鉴定技术从效率、能力与精度方面均难以满足新形势下重大病虫测报要求。针对这一实践需求,以测报灯下害虫图像数据库(约18万张)、田间病虫害图像数据库(约32万张)为基础,构建了基于深度学习方法的病虫害种类特征自动学习、特征融合、识别和位置回归计算框架,并研发了移动式病虫害智能化感知设备和自动识别系统。通过近2年的精确度和实操运行效率检验,该系统在自然状态下对16种灯下常见害虫的识别率为66%～90%,对38种田间常见病虫害(症状)的识别率为50%～90%。随基础数据库的不断丰富、神经网络深层特征提取的不断完善,该系统有望进一步提高识别准确率,从而真正实现田间病虫害识别自动化、智能化和高效率。China has a wide coverage and high dispersion of crops with many diseases and insect pests of which the regional occurrence is complicated. Traditional manual identification techniques are difficult to meet the requirements in terms of efficiency, capability and accuracy. In response to the practical need, based on pests image database under light trapping(about 180,000 pcs) and the field disease and insect pests image database(about 320,000 pcs), a intelligent identification system were constructed in process of feature automatic learning, feature fusion, classification and position regression calculation framework of pests species based on deep learning method, and a mobile pests intelligent sensing device along with an automatic identification system were developed. Through accuracy and practical operation efficiency test in the past two years, the recognition rates of 16 kinds of common pests in light traps under natural condition ranged from 66% to 90%, and the recognition rates of 38 pests symptoms in the field ranged from 50% to 90%.With the continuous enrichment of basic database and the continuous improvement of the deep neural network, the system was expected to further improve the recognition accuracy, thus realizing the automation, intelligence and high efficiency of crop pests identification in field.

关 键 词：深度学习 病虫害 检测 识别 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]