Research On Identification Technique Of Grape Leaf Diseases Based On Deep Learning

Grape is one of the most important cash crops in China,which has made greatly contributed to the development of fruit and wine industries.However,grapes are susceptible to diseases when growing,which leads to a significant decrease in yield and quality.Diseases of grapes are mainly distributed on the leaves,and timely identification and control of those at an early stage of onset can effectively improve the yield and quality of grapes.Therefore,It is meaningful to research a reliable methods for identifying grape leaf diseases to assist farmers in disease control.The early identification of grape leaf diseases was mainly through visual observation,which was subjective and the accuracy of the identification can not meet the requirements.In recent years,deep learning has achieved excellent performance in the field of computer vision with its advantages of fast speed and high accuracy,providing new methods for the identification of crop diseases.Therefore,this paper takes grape leaf diseases as the research target and conducts research on 3D modelling techniques,convolutional neural networks,Transformer and knowledge distillation to construct a dataset and a grape leaf disease recognition model,with the main work as follows.(1)Dataset construction.The grape leaf disease data set Grape A was constructed by collecting four types of grape leaves from the experimental site and photographing them in a natural environment and in the laboratory.To address the problem of insufficient training data and the tendency to overfit the model while training,a method for expanding the grape leaf disease data was proposed.The method converts the grape leaf images taken in the laboratory environment into a 3D stereo model by using 3D modelling techniques.Then Scripted in Python to control the flip of the stereoscopic model around 4 rotation centers and saving 3D image of the it at each angle in real time.Finally,the 3D images were expanded into the dataset for model training.(2)Research on the identification techniques of grape leaf diseases based on improved CCT.In view of the limited number of images of diseased grape leaves,the distribution range and feature details of the disease are many,a method of grape leaf disease recognition,Dens CT,is proposed.This method is based on the CCT model and modifies the single convolutional module in the original model into a multi-convolutional feature extraction module.Besides,the single scale feature extraction unit of 3*3 in new module was improved into a multi-scale feature extraction unit combining 3*3 and 2*3*3.The experimental results showed that the recognition accuracy of Dens CT was 93.92%,which was 6.08% higher than that before the improvement,which verifies the superiority of this method in the task of grape leaf disease recognition.(3)Lightweight grape leaf disease identification techniques based on Mobile Net V3.In order to reduce the space requirements of the model and apply it to a practical environment.A lightweight method for identifying grape leaf diseases based on Mobile Net V3 is proposed.The method modifies the network structure of Mobile Net V3 through CA to enhance the network’s focus on locally important information.Knowledge distillation is also applied to the construction of the model,with the Dens CT network as the teacher network and the modified Mobile Net V3 model as the student network for supervised training.The experimental results show that the improved CA-Mobilnet V3 obtains 93.24% recognition accuracy and its performance approximates that of the teacher network with a size of only 2.78 M,which verifies the superiority of the model.