Research On Rice Diseases Identification Based On Deep Convolutional Neural Network

Rice is one of the most important food crops in China.Rice disease is the main reason for the decrease of rice yield.At present,there are more than one hundred kinds of diseases,among which more than ten are serious.And the prerequisite for doing well in the work of measurement and prevention is correct identification.The application of computer vision in rice disease identification will greatly enhance the accuracy of identification,which is the development direction of agricultural information.In this paper,aiming at the problem of image classification of rice diseases,and for the problems of less training data,low recognition rate and poor convergence in the process of image classification of rice leaf diseases,the convolution neural network model is trained to classify the sample images,so as to achieve the purpose of recognition.In the selection of research objects,four kinds of common disease samples were collected and used as data input.In terms of model algorithm,it is determined to use keras deep learning platform.Among them,the number of volume layers in the network structure of the model is 5,the maxpooling layer is 5 and the full connection layer is 3.The training set consists of four categories,each of which contains 400 sample images,totaling 1600;the corresponding test set has the same composition,totaling 640.Through imagedatagenerator(),the training set and test set are enhanced in different degrees to expand the number of the two kinds of data sets;at the same time,combined with dropout method,the impact of over fitting is reduced.In view of the multi parameters in the model,many comparative experiments are carried out,and the optimal value is determined according to the actual situation.At the same time,according to the characteristics of small samples,we use the transfer learning technology,and use the mainstream vgg-16 model in this study,using two typical methods: feature extraction and fine-tuning,so as to achieve the ideal recognition effect.The experimental results show that the final effect of the model is gradually improved along with the improvement of the method.Among them,the convergence of the model obtained by using transfer learning is more than that of the self-designed model;in terms of training accuracy,both of them are roughly the same,both of them are more than 90%,which meet the basic requirements of disease identification,while the accuracy of verification in transfer learning is significantly better than that of the self-designed model.It provides effective technical support for rice disease identification and control.