Research On Maize Leaf Diseases Identification Method Based On Deep-Transfer Learning

Pests and diseases wipe out nearly half of crop yields worldwide,and the sown area of grain crops in China shows a slight decline.Meanwhile,a change of crop cultivation environment,the variation of disease-causing pathogens of maize and the abuse of pesticides results in the reduction of maize production together with bigger gap between production and demand.Therefore,timely and effective treatment of maize leaf diseases has become a top priority.Infection can occur on all parts of maize at any time during the growing season,but the phenotypic characteristics of leaves can be important identifying criteria.Traditional identification methods are labor-intensive,dependent on the knowledge and experience of identification experts,strongly subjective,and sometimes require the auxiliary diagnosis of laboratory equipment,thus prolonging the diagnosis cycle and delaying the time of disease disposal.In deep learning(DL),convolutional neural network(CNN)has a good performance in extracting graphic features.Meanwhile,in the task of identifying small sample crop disease data sets,transfer learning(TL)shows excellent performance and has become the focus of domestic and foreign experts and scholars.With the advent of agricultural information age,mobile agricultural equipment with low computing power has become a development trend.Deep-transfer learning,which combines DL with TL,can maintain the accuracy,objectivity and immediacy and balance specification and performance through the transfer of lightweight network.Based on deep-transfer learning,this paper implemented the identification of maize leaf diseases and discussed the feasibility of migrating CNN model to mobile devices.The main research work includes:(1)Analysis of the performance of different deep CNN models and lightweight CNN models on the multi-crop disease dataset Plant Village.By setting hyperparameters,online and offline data enhancement,Dropout and L2 regularization,reasonable selection of activation function(Re LU)and optimizer(Adam),the model identification effect has been optimized.An average identification accuracy of Res Net and Mobile Net reached 99.48% and 98.69%,respectively,indicating suitability for transfer learning on local maize leaf disease dataset.The number of parameters and model size of the two network structures are smaller than all the other models,indicating suitability for carrying on mobile terminal devices.(2)Comparison of the identification performance of the CNN pre-trained models in two different source domains(Image Net and Plant Village)into data set of maize leaf diseases.The pre-trained models of Res Net and Mobile Net on Image Net and Plant Village are migrated to the local self-built maize disease datasets respectively,and the training are conducted in the way of layer-freezing and fine-tuning.The Plant Village pre-trained model,which has more similarity with the target domain,achieves higher identification accuracy.Among the 8 TL models obtained,the Mobile Net fine-tuning model based on the Plant Village dataset achieved the highest average identification accuracy of 99.11%.(3)Influences of source domain dataset on the training characteristics of the model in the process of maize leaf disease identification based on deep-transfer learning.After fine-tuning the parameters of the model,not only the training efficiency is improved,but also its average identification accuracy is improved: Res Net and Mobile Net models based on Image Net improved by 2% and 0.84%,respectively,while the Res Net and Mobile Net models based on Plant Village improved by 2.7% and 2.2%,respectively.