Research On Fruit And Vegetable Disease Identification Based On One-shot Learning

Plant disease identification based on deep learning is a hot topic in research of modern agriculture.In fact,the problems such as the low incidence,long period and the high cost of plant disease data collection result the less sample collection.Deep learning neural network model training relies on large data set and the plant disease datasets are too small to affect the application of the model in the field of plant disease.Transfer learning aims to extract the knowledge and features from one or more large datasets,and applies the knowledge and features to a new small dataset.In the meanwhile,data augmentation technology can expand the dataset by translating,rotating and distorting images to improve the recognition accuracy.However,transfer learning will acquire a low accuracy for the identification of the single sample data.To tackle the low recognition rate of model due to the small number of data samples,this paper proposed the combination of one-shot learning for fruit and vegetable disease identification.The experiment used the public PlantVillage,and analyzed the distribution of the public dataset.And it studies the 8 kinds of fruit and vegetable diseases which given only few examples from each.The focal loss function is used instead of the mean square error to train the fruit and vegetable disease classifier based on relation network.During training,the model put focus on the hard examples to improve the accuracy of plant disease identification by adjusting the hyper-parameters.The method achieved 91.64%accuracy for 3-way 1-shot classification task,which improves about 1.41%accuracy than original model.The method improves about 4.69%for 5-way,1-shot.Besides,compared with Matching Network and transfer learning,this method can effectively solve the problem of fruit and vegetable disease identification training on the small dataset and accurately identify plant disease.In addition,the model is deployed to the server to identify the new class with fewer samples in the real-time,which promote the development of modern agricultural economy.