Research On Image Classification Methods Of Crop Diseases Based On Meta-Learning

Crop identification and diagnosis is the premise of prevention and control,which has practical significance to ensure high yield,stable yield and high quality of crops.In this paper,the image of crop diseases(rice blast,wheat scab,etc.)is taken as the research object,and the meta learning method of image classification of crop diseases is mainly researched.The image classification model of crop diseases based on meta learning and transfer learning is designed,and the performance differences between them are compared on different scale datasets.The main work is as follows:(1)A network model of crop disease image classification based on meta-learning is designed.Based on MAML,Reptile and improved MAML methods,combined with the ResNeXt101 network,three crop disease image classification framework models were constructed;through experiments,various key hyperparameters and structural optimization settings of the model are explored,and the Group Normalization strategy is used to replace the original batch Batch Normalization;the source domain data set PlantVillage is used as the training set,and the target domain IDADP small batch data set is used as the test set.The experimental results show that the improved MAML method combined with ResNeXt101 model is the best,with an accuracy of 83.33% on 300 labeled samples.(2)A network model of crop disease image classification based on transfer learning is constructed.AlexNet,VGG-16,VGG-19,Mobile Net＿V3＿Small and ResNeXt101 were transfered based on Fine-tune technology;for comparison,the hyperparameter settings of the transfer learning model are consistent with the meta-learning model;The IDADP data set is divided into training set and test set of different sizes.The experimental results show that the accuracy of the transferred ResNeXt101 model is 95.67% after training on 3000 samples,while the accuracy of the model is 49.33% when the training set is reduced to 300 samples.(3)The performance of crop disease image classification model was analyzed based on comprehensive evaluation index.Through macro average precision rate,macro average recall rate,macro average F1 and execution time and other indicators,comprehensively compare the execution efficiency and generalization ability of the model.The results show that in the case of limited samples,the meta-learning method has higher execution efficiency and shorter classification time than the transfer learning method.Moreover,it is not limited to image classification under a single label space,showing stronger generalization ability.The crop disease image meta-learning classification model constructed in this paper realizes the rapid,reliable and effective identification of crop diseases under the condition of limited labeled samples,and provides technical support for the diagnosis and prevention of crop diseases.