Classification And Identification Of Crop Disease Images Based On Deep Convolutional Neural Network

China is a large agricultural production country with many types of crops.It is also a country with frequent and recurring diseases.The occurrence of diseases has the characteristics of many categories,wide range,high frequency and heavy degree,which has a great impact on the national economy,agricultural production,especially food security.For example,only in 2019,the incidence of major diseases in China reached 4.3 billion times per mu,and potato late blight and maize leaf spot occurred moderately in some regions of China,causing certain economic losses to farmers.In the classification of crop diseases,only the classification of diseases is often identified,and the severity of the disease is not carefully classified.In this paper,the data set used contains 10 different crops such as potato,corn,tomato,etc.This data set is divided into 60 categories according to the types and disease conditions(general and serious)of crop diseases.Based on deep convolutional neural network and transfer learning,we carry out crop image classification research.The main work is as follows:(1)Building a deep convolutional neural network,which is composed of the input layer,four convolutional layers,four pooling layers,one fully connected layer and the output layer.Training it on crop disease datasets.On the basis of the original data set,different data sets were set up for comparison experiments.During the training process,data enhancement,Dropout,Adam optimization algorithm and other technologies were adopted.The model trained in less epochs of 100 can achieve the average classification accuracy of 83.75%.(2)To solve the problem that the classification is a tough task due to unbalanced samples,the deep networks of VGG19,inception V3,Res Net50V2 and Mobile Net V2 are constructed via transfer learning.Subsequently,the transfer training on data set is completed and applied into crop diseases.Finally,according to the full connection layer,optimizing the loss function and fine-tuning the pre-training models,our model which has strong generalization ability is proposed,where the feature extraction has been visualized.The experimental results show that our transfer learning model can be achieved rapid convergence within 50 epochs,and its accuracy can reach 96.26% within 200 epochs for the classification of crop diseases,which is better than that of the other state-of-the-arts.In summary,our deep convolutional neural networks is lightweight and can classify crop leaf diseases effectively.Using transfer learning and fine-tuning pre-training VGG19 models,the accuracy rate and robust of our model can be greatly improved.A fine classification of the disease status of crop diseases is significant to understanding the degree of damage to crops,proper fertilizing amount,reducing pesticide residues.