Rice Leaf Disease Classification Based On Convolutional Neural Network

Rice is one of the important food crops in China and the world,and the safe production of rice is a strategic issue directly related to social stability as well as national security.Due to various factors such as artificial farming system as well as climate,environment and soil,which directly or indirectly affect the normal growth of rice,various disease problems have emerged,which seriously affect the yield and quality of rice.Currently,most researchers use traditional image techniques and machine learning to study one disease or several diseases of a certain crop,and these techniques are highly dependent on specific disease characteristics.Although certain achievements have been made in the direction of rice disease classification,the classification models obtained also have certain limitations,which can be solved to some extent by using convolutional neural networks,which can The use of convolutional neural networks can solve these limitations to some extent by training a large number of sample data to achieve automatic extraction of disease feature information.To address the above problems,this paper uses convolutional neural networks to study several diseases commonly found in rice leaves,and the main research contents are as follows:(1)In this thesis,the rice leaf disease dataset was obtained through Kaggle platform and Internet search,and all the obtained rice leaf disease images were classified according to their different background complexity into Data1 dataset with simple background and Data2 dataset with complex background,and the two datasets were subjected to data enhancement process.(2)Five classical convolutional neural networks,VGG16,Goog LeNet,MobileNet V2,ResNet34,and ResNet50,were used to conduct classification experiments on the Data1 dataset in simple background,and the results showed that ResNet50 had the highest classification accuracy with the same input rice leaf disease dataset,and the model was used to Experiments using the model for Data2 dataset with complex background,the classification accuracy still needs to be improved,which is because the complex background of rice interferes with the experiment,so the model needs to be improved.(3)Based on the ResNet50 model,the ResNet50 model was improved by null convolution,cosine annealing to attenuate the learning rate and incorporating the CBAM attention mechanism,and then an improved ResNet50 rice leaf disease classification model was constructed.The parameters of the ResNet50 model were set by experimental investigation,and then the feasibility of cosine annealing decay learning rate and the improved model incorporating CBAM attention mechanism were verified by experimental comparison with Goog LeNet,MobileNet V2,and CA＿ResNet50,and finally the classification accuracy of the improved model on the test set was obtained to be 97.9%,which is higher than other models.Experiments using the improved model on Data1 dataset with simple background also showed some improvement in classification accuracy.Finally,the model was used in combination with Py Qt to design a rice leaf disease classification system,which was used to classify rice leaf diseases.