| Accurate identification of agricultural pest species in real time is an important prerequisite for effective pest control.At present,the diagnosis methods of crop pests in China mainly rely on artificial identification,subjective factors are large,real-time is poor,and there are fewer insect taxonomists,and farmers' professional knowledge is relatively lacking.In the case of the decreasing number of grassroots plant protection personnel,it is urgent to have convenient Crop pest intelligent identification tool.With the popularization of smart phones and the successful application of deep learning in image recognition,it has become a hot spot for many scholars to use mobile phones to capture pest images and identify them.Based on 34 common crop pests,this paper establishes an intelligent recognition system for agricultural pest images based on deep learning,including mobile client,elastic compute service and deep learning algorithm,which realizes efficient,real-time and accurate identification of pest images.The main research contents and results are as follows:(1)This article developed a client application based on the Andriod platform.Users can log in by registering an account,SMS verification or QQ verification.When the user encounters an unknown pest in the agricultural production,the client application software can be used to capture or select the pest image in the album,and after previewing,cropping,etc.,it is uploaded to the server for identification diagnosis.The method synchronizes the recognition result to the map interface to learn about the distribution of pests around and across the country.Users can also learn and understand crop pests by means of inquiry,and learn the chemical control and biological control measures.Users can view identification records in the personal center,and can also conduct real-time online question and answer through remote expert diagnosis function for uncertain results.(2)This paper builds a cloud server based on Windows Server system.Receive and process data such as login information,user registration information,agricultural pest images,GPS information,and expert diagnostic information by writing Servlet business logic;realize the design and connection of database management system through MySQL and JDBC;Decoupling between business code and database operations through DAO mode;using the connection pool to optimize the database connection;the Caffe deep learning framework is built and reconstructed on the microcomputer,and the project is packaged into a War package and deployed on the cloud server.(3)Research on agricultural pest identification algorithm based on deep learning.Based on 34 major crop pests and more than 10,000 images,the identification model of agricultural pests was obtained by fine-tuning the convolutional neural network structure on the widely used CaffeNet,VGGNet and ResNet.By testing and comparing a large number of pest images,the 101-layer ResNet model has great recognition effect,and the average recognition accuracy can reach 93.5%,which has good robustness and generalization ability.(4)The call of the deep learning model is implemented using the JNI native method interface and the DLL dynamic link library.The identification of agricultural pest images is achieved by adding model files,address files of the mean files,weight files and tag files in the Native native method.After testing,the average recognition response speed of agricultural pests in CPU mode is about 1.1 seconds,and the average recognition response time in GPU mode is about 0.67 seconds,and the deviation is within 200 ms,which satisfies the needs of practical applications.The agricultural pest image intelligent identification system studied in this paper can view pest information online,identify the pictures of pests uploaded by users in real time,and guide farmers to carry out corrective activities correctly.It has a good application prospect and can be widely used in the identification and diagnosis of crop pests in the field. |
PDF Full Text Request