Research On The Methods Of Grape Leaf Disease Identification Based On Deep Learning

Grape cultivation has a long history in our country,which not only has good nutritional value,but also widely used in wine industry,yet its large-scale cultivation will also lead to leaf diseases.If not timely control it will affect the quality of fruit,leading to a reduction in yield and other negative consequences.In order to deal with the occurrence of disease spraying pesticide in advance is usually used to prevent the occurrence of disease,but in the process of it,there is the problem of excessive pesticide application,which not only pollutes the environment,but also increases the cost of planting.Traditional disease recognition technology is not ideal for crop images with more distractors and objects similar in color and shape.The disease recognition rate is low,and there are many omissions and errors.Traditional segmentation methods,such as threshold segmentation,K-means clustering,watershed algorithm,etc.,are prone to cause oversegmentation,error segmentation,noise sensitivity and poor segmentation performance.Therefore,in order to better solve the problem of pesticide pollution and reduce pesticide abuse,it is of great significance to study more accurate and efficient grape leaf disease recognition algorithm.At present,the wide application of deep learning in agriculture provides a theoretical basis for nondestructive testing of grape leaf diseases.The image recognition method based on deep learning realized the accurate recognition of leaf disease types,and the image segmentation method based on deep learning realized the accurate segmentation of leaf disease spots.Finally,the application of deep learning in grape leaf disease recognition realized the accurate determination of the degree of leaf disease,and provided better technical support for the early prevention and control of grape leaf disease.In this paper,the recognition and classification of grape leaf disease images will be carried out based on convolutional neural network theory and image segmentation algorithm,so as to provide a new method for the early intelligent diagnosis of grape leaf disease.The main research work is as follows:(1)Considering that the number of images collected and collated in the experiment is limited,the size of the shot images is inconsistent,this paper uses data enhancement techniques such as rotation,translation,miscut and scaling to expand the original data set.At the same time,in view of the problem of different image sizes of grape disease data sets,the images were scaled before being inputing into the model to keep the image sizes in the data set consistent.Finally,the data sets were manually labeled and made into standard data sets to prepare for the subsequent experimental research.(2)Aiming at the problem that the existing grape leaf disease spots are small and dense,which leads to the low accuracy of small target disease recognition.Target recognition is easy to be missed and for images with more interferors,a YOLOv5 s model based on attention mechanism(CB-YOLOv5s)is proposed.In this model,CBAM attention mechanism was introduced into YOLOv5 s.This attention mechanism was embedded into the convolutional layer of backbone network to improve the feature extraction ability of small lesions and the detection effect of easily ignored leaf diseases.By using EIo U Loss as the network regression loss function in the model,the convergence speed of the model is accelerated and the recognition accuracy is improved.After the improvement,the m AP value is 92.6.(3)Aiming at the problem that the size and shape of grape leaf disease spots vary with the occurrence time and degree,resulting in poor segmentation ability of large or small disease spots in the image due to different target shapes and sizes,a U-Net model based on Inception multi-scale module(IN-Unet)was proposed.It was used for image segmentation of grape leaf disease.The multi-scale feature extraction module is added to each subsampling layer,and the fusion of multi-scale features is realized through concat operation,which solves the poor segmentation effect caused by inconsistent target sizes and improves the model performance.The improved MPA value is 84.79.The relative area ratio of the spot area and the leaf area was calculated to realize the disease classification diagnosis.