Research On Methods For Grape Leaf Disease Recognition Based On Deep Learning

With the rapid development of Chinese wine industry,the production and sales volume of wine are steady growing.The planting patterns of wine grape and table-grape are the large-scale,high-density and clustered which have brought great challenges to the prevention and treatment of infectious diseases.In the past,people diagnosed diseases via the infection test or their experience to observe the characteristics of disease,but it takes long cycle and subjectively.It is easy to cause pesticide residues and disease resistance due to spraying pesticides blindly.With the advancement of technology,artificial intelligence,computer vision and image processing technology are increasingly used in agriculture,especially in the field of crop disease identification.When grape plants infected,the physiological structure and morphological characteristics will change,such as deformation,fading,decay and so on,so the image of grape leaf disease can be used as the object.It is significance that identify disease species with the images of infected grape leaves based on computer vision technology and deep neural network for the quality of agricultural products and agricultural ecological environment security.This paper outlines the classification methods of 6 kinds of common diseases in grape leaves under natural illumination,including the following contents:(1)The collected grape leaf images were enhanced without changing the properties and categories of the research object,using the methods of mirror flipping,cropping randomly,adding color disturbance and noise to increase the number of additional copies in order to solve the problem of inadequate samples.3 kinds of datasets were made according to different recognition methods that are Dataset A,Dataset B and Dataset C.(2)Semi-automatic identification method of grape diseases: The experimental samples were the intercepted rectangular region of diseased spots which were taken from original leaf image containing the complex background manually.Firstly,28 kinds of characteristics were extracted according to the differences of 6 kinds of grape diseases in color and texture,and 20 features were selected through feature screening.The effective feature of the samples was realized by PCA(Principal Component Analysis),and 22 traditional classifiers were tested respectively.Then 5 kinds of features were used to classification according to the contribution rate of prediction results.In addition,in order to study the effect of the deep learning method on the spot area.The experimental results show that the Bagged Trees in the traditional classification method has better recognition effect on the grape disease image based on color and texture feature,and the average recognition accuracy of 6 diseases is 86.67%.This method showed high recognition rate and robustness especially for powdery mildew,and the recognition rate is 92.94%.(3)Automatic identification of grape diseases: The faster R-CNN model was used in this paper to detect the leaf area in the images that are complete diseased leaf samples containing complex backgrounds.The disease spots are detected in the previous detected leaf area,then outlining the leaf area with the outer rectangular frames,and the rectangular images are finally used to identify the disease type by CNNs.Faster R-CNN is still used to detect disease spots.This strategy can effectively avoid the error detection and false recognition caused by complex background factors and ineffective regional disturbances.In this paper,the Bagged Trees of the traditional machine vision classification models is tested and compared with the identification method of the artificial interception of the disease spot samples in the previous paper.In addition,in order to select the optimal network in a variety of convolution neural networks,this paper tests the AlexNet,VGG-16,VGG-19,GoogLeNet,ResNet50 networks.The experimental results show that the classification performance of the samples with complex background is not significantly improved by the traditional classification technique,and the correct recognition rate is 73.99%.In 5 kinds of network models,VGG-16 showed excellent recognition performance,and the correct recognition rate of 6 kinds of grape diseases was 94.48%.The method has better recognition effect on anthracnose,Botrytis cinerea,brown spot,powdery mildew and black pox,and the correct recognition rate is higher than 90%,among which the identification rate of powdery mildew is up to 99.62%.