Identification Of Grape Leaf Diseases Based On Transfer Learning And Software Design

In recent years,With the continuous development of digital farming technology,the agricultural expert system based on machine learning has been continuously improved,but also enriched a large number of effective expert knowledge base.However,the rapid and effective automatic monitoring of crop diseases using machine vision technology and deep learning theory is still in the preliminary stage of research.This paper takes the image of grape leaf disease as the research object,fully explores the characteristics of grape leaf shape,color,disease spot and so on when the grape plant comes on,uses the deep learning theory and the migration learning technology to carry on the modeling research,and deploys the optimized model to the mobile terminal for rapid application,which greatly facilitates the identification and prevention of grape disease by the growers and the quality of agricultural products It is of great practical significance to improve the yield,apply medicine reasonably and protect the safety of agricultural ecological environment.Based on the AI Challenger 2018 crop disease data set and four kinds of grape leaf images acquired by two ways of field collection,this paper studies the identification of grape leaf disease.The main research contents are as follows:(1)From the perspective of the current situation of crop disease research,this paper reviews the methods and theories that some researchers at home and abroad use machine vision technology to identify and diagnose crop disease,discusses the current development of digital farming technology in China and the development of crop disease identification software in mobile applications,and puts forward the main objectives of this paper.(2)Based on the analysis and summary of the characteristics of some existing deep learning network models,in order to improve the recognition accuracy of the deep learning network model and ensure the portability of mobile terminals,this paper selects the mobilenetv2 network model for migration learning.In addition to the output layer parameters,a layer of pooling layer and full connection layer are added and the output is activated through the softmax function.The results are as follows: Good results.The model is optimized by adjusting the training parameters and super parameters of the migration model,so that the final model in the test set reaches 96.6% accuracy and the model volume is about 21 MB.Compared with the migration of squeezenet,shufflenet and mobilenetv2 models,considering both the storage space and prediction accuracy of the model,the mobilenetv2 migration model can take into account the advantages of reducing the amount of parameters and calculation,and is more suitable for the migration of mobile terminals.(3)In order to facilitate the rapid and automatic identification of grape diseases by ordinary farmers,the deep learning model constructed in this paper is transplanted to Android mobile terminal.The software system includes four parts: grape disease knowledge,grape pest knowledge,disease identification and statistical report.Through selecting 50 samples of four grape leaves and 200 samples of them for Android real-time detection,the overall recognition accuracy is 95.5%.The system basically meets the needs of ordinary farmers for off-line rapid automatic recognition of grape diseases.