Research On Grape Leaf Diseases Identification Method Based On Transfer Learning

Grape is one of the most popular fruits with large planting area in China.The stable development of grape industry plays a positive role in China’s food safety and social and economic development.However,in the process of grape planting,diseases have become an important factor restricting the healthy growth of grapes.At present,grape diseases in China mainly include black rot,black measles,leaf spot,downy mildew and phylloxera,among which black rot and downy mildew are particularly serious.Therefore,timely and accurate identification of leaf diseases is greatly significate to timely control grape diseases and to reduce economic losses.In order to realize timely diagnosis and accurate identification of grape leaf diseases and ensure the stable development of grape industry.In this thesis,five kinds of diseased leaves and healthy leaves are taken as the research object,and the identification of grape leaf disease images based on transfer learning and deep transfer learning is mainly studied.It provides theoretical basis and technical support for intelligent,rapid and accurate identification of grape leaf diseases.Research content of this paper as follows.(1)Aiming at the problems of lack of large data sets and low quality of data sets of grape leaf diseases,a grape leaf disease identification approach based on transfer learning and SENe Xt is proposed.Firstly,the approach combines the advantages of the Squeeze and Excitation(SE)network with few parameters,easy integration,fast operation and simplified design of Res Ne Xt network,SENe Xt convolutional neural networks(CNN)is constructed by integrating the SE module into the Res Ne Xt network.Then,the data set of grape leaf disease is expanded online by combining spatial transformation and color distortion data augmentation approaches,and the enhanced data set is constructed.Finally,the SENe Xt model and transfer strategy are used to construct the identification model of grape leaf diseases to avoid the overfitting problem caused by training small sample data set directly by the deep learning model.The experimental results show that the classification performance of SENe Xt disease identification model is better than that of the classical algorithms Res Ne Xt＿50 and Res Net＿50,and the accuracy is improved by 4.69% and5.94%,respectively,compared with Res Ne Xt＿50 and Res Net＿50.In addition,the accuracy of transfer learning approach on SENe Xt model is 97.20%,which is higher than the new learning performance and faster convergence speed.(2)To solve the problem of large parameters and high complexity of SENe Xt network model,a simple and efficient approach of grape leaf disease identification based on deep transfer learning and Mobile Net V3(GLD-DTL)is proposed.In this approach,firstly,knowledge distillation strategy is adopted to train mobilenetv3 student network with Res Ne Xt101 as teacher network,which is improved through model migration strategy,and a lightweight grape leaf diseases recognition network model(GLDR)is further constructed.Then,an image quality detector is designed to reduce the influence of blurred image on the recognition effect of grape disease.Based on this,a data augmentation approach combining inverse filter noise and spatial transform is proposed to expand the diseased grape image,so as to avoid the overfitting problem caused by unbalanced sample distribution.Finally,a three-stage training strategy is used to train the GLDR model to prevent the direct fine-tuning expansion of the network from negatively affecting the backbone network.Results show that the accuracy of GLD-DTL approach in early diagnosis of grape leaf diseases is 99.84%.Compared with VGG19 model,the accuracy of GLDR model is improved by1.26%.In addition,the size of the model is only 30 MB.Therefore,the proposed approach satisfies the requirements of accuracy and size of the model in mobile deployment.(3)In order to solve the problems of complex deployment,slow identification speed and high requirement of network bandwidth,a grape leaf disease identification system with fast inference speed and simple deployment is proposed.In this system,Python and Java are used as the main development language,and the GLDR is used as the background model to develop the application program of grape leaf disease recognition.The system interface is simple,easy to operate,does not rely on communication networks,cloud platform servers for processing,saving bandwidth and transmission time.