Analysis And Research Of Deep Learning Based Rice Disease Identification Method

For centuries,rice has been one of the most important crops in China and its production has received much attention.Nowadays,with a population of 1.4 billion,close to the majority of the country’s population chooses rice as a staple food.However the emergence of a variety of rice diseases has caused huge losses in rice production and also brought great trauma to our agricultural economy.Therefore rice diseases should be detected and confirmed early to make it reduce certain losses,and it is necessary to design a method to identify rice diseases quickly and accurately.The use of traditional manual identification methods to identify rice diseases is not only time-consuming and labor-intensive but also requires considerable expertise to identify,and the accuracy rate is not well guaranteed.With the gradual increase in the incidence of rice diseases,the requirements for traditional identification methods have also increased.In recent years,China’s science and technology in the direction of network models continue to develop,based on the convolutional neural network model of rice disease identification system technology has gradually matured.The use of neural network models for rice disease identification is a fundamental solution to reduce the cost of labor,and the speed is also greatly improved,while the accuracy can also reach a certain level.In this paper,four rice diseases with high incidence,namely,Donggelu disease,white leaf blight,rice blast and brown spot disease,were selected as the main research subjects.The collected data images are processed using pre-processing means and data enhancement such as rotation,cropping and mirroring.The model designed in this paper is based on the VGG model and Res Net model to improve the structure of the classifier,replacing the VGG three fully connected layers with a layer of global average pooling to reduce the total number of model parameters and training time,and subsequently the convolutional,pooling and fully connected layers are designed and parameter adjusted to constitute a relatively complete convolutional neural network model,the.To verify the performance of the designed model,the processed data were input to the model for model training with an accuracy of more than 98%.In addition,parameter discussion and comparison of different models were conducted to show that it is optimal and the improved model can reach 98.68% in accuracy at 800 iterations,learning rate of 0.001 and batch size of 64.Finally,the model is maximized using the Py Qt5 system recognition platform.By inputting the disease images into this platform,the disease type can be obtained very quickly to ensure that the subsequent treatment and protection can be carried out quickly and effectively.