Research On Grape Leaf Disease Detection Based On Convolutional Neural Network

In recent years,agricultural production is gradually developing for the direction of automation.There have been research achievement on machine seeding and automatic irrigation in agricultural production,but the detection method based on machine vision is still under research.At present,automatic detection method of disease based on machine vision is generally carried out under ideal conditions,the research subjects were placed under simple background for testing.In this paper,the grapevine leaves were taken as the research object,and the surface lesions of grape leaves under natural conditions were studied.The machine vision-based method was used to detect the disease automatically.In the monitoring images,we have used the surface information of the blade detected through the computer.The status assessment of growing grapes can be achieved.The detection algorithm of this paper uses the Faster-RCNN in the multi-angle suggestion area to accurately locate the grape leaves in the image and detect the disease on the detected leaf.Compared with the direct detection of the disease in the image,the interference of the background factor to the disease area is eliminated,and the false detection rate is effectively reduced.During the leaf detection,this paper proposes an improved proposals algorithm to generate higher-quality blade candidate regions in an image,which can be used to adapt to grape leaves with various postures under natural conditions.the experimental data show that data show that the algorithm has good adaptability to the detection of grape leaves in natural images.In the experiment,the image of six different weather conditions has been statistically recorded,and the average mAP of the average leaf detection algorithm is 75.52%,which is significantly higher than traditional algorithms.During the detection of the disease,considering that the detector will misjudge the background area in the image as a disease area,this paper proposes two methods for detecting diseases.For the first method,each single leaf detected from one image,or the entire image is masked,is used as the input image for the next level of disease detector.The experimental results show that for the first method,the average mAP for 6 common grape diseases is 66.47%,and the mAP for brown spot and powdery mildew exceeds 70%;for the second method,the average mAP for disease detection is 51.44%.However,the average detection time is 75% less than the first method.Both methods perform better than the direct disease detection method on the original image.In order to visually display the effect of disease detection,this paper unified all the processes of the disease detection into the graphical interface based on Matlab,designed GUI human-computer interaction interface for blade detection,disease detection and other modules.In the process of image acquisition,the tree blackberry was used to control the remote camera,and the online monitoring was realized by reading an image captured in the camera or loading a local image.