Improvement Of Few-shot Classification Method Which Based On Contrastive Learning And Its Application In Grape Leaf Disease Classification

In recent years,the demand for grapes and related products has been increasing,and diseases are the main cause of grape yield reduction,causing huge economic losses to farmers every year.Therefore,timely and accurate identification and classification of diseases is crucial.In view of the problems existing in small sample learning,this paper designs a new network structure and its training method to improve small sample learning,in which a convolutional network is used for feature embedding,and multi-scale sliding pooling is combined to enhance feature extraction.The main structure of the network is the twin network,in order to learn semantics from small sample data through sample comparison.The training method of the network adopts nested hierarchical parameter updates to ensure the convergence stability.Therefore,this paper uses contrastive networks for few-shot learning(CNFS)to recognize and classify grape leaf diseases.The experiment selected healthy grape leaves and three kinds of diseased grape leaves as datasets.After the model is trained,VGG,Resnet and the improved CNFS model were used to compare the classification problems on small sample data sets,and the identification rate reached about 96%.The experimental results show that the CNFS model have a good recognition effect on grape leaf diseases,and can be effectively applied to agricultural production.