Research On Grape Disease Identification Based On Deep Learnin

China has become the world’s largest grape consumer and producer and the grape industry has helped many places become rich.Accurate and rapid identification of grape disease types and disease severity is an important guarantee for increasing production and income because diseases affect grape quality and fruit yield.In this paper,we focus on the image classification and semantic segmentation of diseases on grape leaf based on deep learning.It solves the misjudgment caused by growers relying on experience for a long time to identify,and provides technical support for automatic,efficient and accurate diagnosis of grape leaf diseases.The major research contents and results of this paper as follows:(1)Build a grape disease dataset and complete classification and identification.Enlarge the dataset using data augmentation techniques such as geometric transformations and color transformations and established datasets of three common grape leaf diseases and healthy grape leaves.Combined with the characteristics of Res Net residual structure that can make features fully circulate and Swin Tansformer with fewer parameters and fast operation,Swin Transformer network with residual structure was constructed,and finally the classification accuracy of grape disease data set reached 97.138%.(2)In order to realize the image segmentation of grape disease,a semantic segmentation network based on Swin transformer is constructed.Incorporating the Swin Transformer and attention mechanism into the encoder and decoder of the U-Net network respectively.Experiments show that using channel attention on small feature maps and spatial attention on larger feature maps can improve the performance of the encoder with fewer parameters.The results of MIo U and MAP were 87.82% and97.60%.(3)Design and realization of grape leaf disease identification system.The mobile terminal is deployed based on the We Chat applet,which can easily complete the blade collection work.The management platform adopts the Vue framework and is written in HTML,CSS and Javascript,which can complete the management of the diseased area and the update of the identification model.The back-end is based on the Spring framework.The code base on Java and the deployment of SW-Unet on the server is completed through Flask.