Plant Disease Recognition And Severity Assessment In Mobile Phone

The relationship between human beings and crops is closely connected on the earth.Crops provide a lot of use values for human beings,such as: edible,medicinal and so on.At the same time,it plays a role in protecting our earth from air pollution.In agricultural production and research,disease has become one of the most important factors affecting plant growth,high yield and income.Doing a good job of plant disease identification can better guarantee plant growth and timely get treatment measures for the disease plants.Traditional identification and detection of crop diseases mainly rely on experts with rich plant disease identification to identify diseases in crop fields.This artificial field identification method needs high knowledge reserve for experts in plant disease identification,which wastes manpower,has low recognition efficiency and complex working process.It is hard to avoid human subjectivity in disease classification and to grasp the accuracy.With the development of image recognition technology,people began to apply pattern recognition,computer vision and other computer technologies to plant species recognition.In modern times,the main research aspect of plant disease recognition is used in the computer end.The main research object of this paper is mainly in the hand end,and the development of disease recognition app in the mobile phone end makes the disease recognition operation more convenient and convenient for farmers On site use.The main research work of this paper is as follows:(1)Detection and identification of plant diseasesThis paper studies the diseased leaves of plantvillage data set and natural data set.In order to realize the detection of diseased leaves on mobile phone,mobilenet and perception V3 two kinds of light convolution neural networks are used for migration learning,and two kinds of recognition network models are obtained.The two network models are transplanted to Android mobile phone respectively,balancing recognition accuracy,operation speed and network size,and selecting the optimal model.The results show that the average recognition accuracy(ACC)of mobilenet and perception V3 on the plantvillage dataset(38 categories and 26 diseases)is 95.02%and 95.62%.In the self built dataset,the average recognition accuracy(ACC)of mobilenet and perception V3 is 87.50% and 88.06%.The overall recognition accuracy of perception V3 is slightly higher,but mobilenet has a better balance in all categories;perception in model size The model size of V3 is 87.5m,and the model size of mobilenet is 17.1m,about 5 times of the latter;when the two models are transplanted to the mobile terminal,the memory occupied by mobilenet and perception V3 app is 21.5m and 125 m respectively;in terms of the recognition time of a single picture on the mobile terminal,perception The average calculation time of V3 is about 174 ms,the average calculation time of mobilenet is about 134 ms,and the average calculation time of the latter is 40 ms faster than the former;compared with perception V3,mobilenet on the mobile terminal takes up less memory and faster calculation time.It shows that mobilenet is more suitable for the application of plant disease identification on mobile phone.(2)Detection of plant disease degreeIn order to estimate the severity of leaf disease,it is usually necessary to segment the leaf and disease area and calculate the area of both.However,there are few depth models suitable for semantic segmentation on mobile phones.In this paper,leaf and disease area prediction based on disease spot detection is used.This method uses mobilenet＿The SSD target detection algorithm detects the leaves and the diseased spots,and establishes the regression model between the boundary frame size of the detected leaves and the diseased spots and the manually segmented areas of the leaves and diseased spots.MobileNet＿SSD can be deployed on mobile phones,so through the target detection algorithm can predict the area of plant leaves and disease spots,so as to estimate the severity of leaf diseases.The experiment shows that the model has a high degree of fitting between the predicted area and the actual area,and can well predict the disease degree of plant leaves.