Research On Method And System Design Of Grape Downy Mildew Grading Based On Machine Vision

Grape downy mildew is one of the most harmful grape diseases,which mainly affects the leaves.If it is not prevented timely,the whole grape orchard will be infected with the disease,greatly reducing the quality and yield of grapes,resulting in serious economic losses.At present,it is difficult to accurately identify the level of this disease,resulting in unreasonable application of pesticides,environmental pollution and waste of liquid medicine.In order to reduce pesticide use,improve efficacy and protect agricultural ecological environment,it is particularly important to study the grading detection algorithm and grading detection system of grape downy mildew.In this study,grape downy mildew disease was identified by machine vision combined with deep learning.Aiming at the grading and evaluation of grape downy mildew with complex background in natural environment,this study proposed a grading and detection algorithm of grape downy mildew based on semantic segmentation combined with K-Means clustering and random forest,which realized the rapid grading and detection of grape downy mildew and provided algorithm support for the grading and evaluation of other similar diseases.Based on the above methods,this study designed a grape downy mildew grading detection system.Android smart phones were used as acquisition and display terminals,and the complex calculation process was put on the local computer for calculation.The main research contents of this paper are as follows:(1)Grape leaf extraction model under complex background based on deep learning.The data set of orchard grape leaf with complex background in natural environment was constructed.Considering the complexity of actual orchard environment,HRNet v2+OCR algorithm was used to construct grape leaf extraction model.The accuracy and average intersection ratio of different semantic segmentation networks were compared,and the segmentation performances of different segmentation algorithms were analyzed.The results showed that HRNet v2+OCR algorithm could better segment grape leaf in actual orchard environment than other algorithms,and its recognition accuracy was improved compared with other common semantic segmentation networks.Based on W48 trunk network,the accuracy of this model was 98.45%,and the average crossover ratio was97.23%.(2)Grape downy mildew spot segmentation algorithm based on machine learning.An algorithm for disease spot segmentation was proposed based on K-Means and random forest.The images that could not be directly labeled with disease spot regions were learned by using split images,and only a few split images were used to complete spot training.At the same time,combining with semantic segmentation algorithm,a pixel size transformation strategy was introduced to make the segmentation algorithm have more information and higher resolution.The grading accuracies of downy mildew on positive,negative and both sides of grape leaves by combining K-Means clustering and random forest were 52.59%,73.08% and 63.32%,respectively.The grading accuracy of downy mildew on positive,negative and both sides of grape leaves when the disease grade error was less than or equal to 2 were 88.67%,96.97% and 92.98%,respectively.(3)Terminal deployment of rapid grading detection algorithm for grape downy mildew.Based on the above methods,a grape downy mildew grading detection system was constructed,which was mainly composed of an Android application terminal,a cloud server and a deep learning computing terminal.The client was developed based on Android and designed in C/S mode.It mainly realized the collection of grape downy mildew image data,location information collection,image data uploading to the server and the display of the results of operation.The Cloud server adopted Spring Cloud framework of distributed micro-service for development,and adopted Cloud platform for project deployment,mainly realized the development of Cloud server and terminal devices based on HTTP and Web Socket,communication between Cloud server and deep learning computing terminal based on GRPC,task scheduling and persistence,etc.The deep learning computing terminal was developed by Python and mainly realized the communication with the cloud server and the deployment of grape downy mildew grading detection algorithm.Validation tests showed that the system could help users identify the grade of grape downy mildew,explain the corresponding prevention and control measures,and realize the grading detection of grape downy mildew in the orchard environment.