Identification Of Tea Leaf’s Diseases In Natural Scene Based On Convolutional Neural Network

Tea is an important economic crop in my country.During its growth,tea is often infected with various diseases such as tea bud blight,tea leaf blight,tea red leaf spot and tea algae leaf spot.Accurate and rapid identification of tea leaf’s diseases is beneficial to their prevention and control,and improve the yield and quality of tea.With the development of computer technology,certain achievements have been made in using machine learning to identify crop diseases.However,due to the complex backgrounds,relatively small disease spots,and similar colors and textures in tea leaf images taken in natural scenes,the accuracy of using existing algorithms to identify tea leaf’s diseases in natural scenes is not high.In order to improve the accuracy of tea leaf’s disease identification in natural scenes,this thesis preprocesses the images based on the characteristics of tea leaf’s disease images in natural scenes and improves the convolutional neural network model.The main work done in this thesis is as follows:(1)Two image data sets of diseased tea leaves were constructed.Using Canon EOS 80 D SLR camera to take images of diseased tea leaves in Anhui Agricultural University and Tianjingshan Tea Garden in Anhui Province for many times,construct two diseased tea leaf image data sets.In addition,in order to better train the tea disease identification model,the training sample images in the data set are manually labeled and augmented.(2)A tea leaf’s disease identification algorithm based on improved convolutional neural network model and SVM is proposed.The algorithm first uses a multiscale feature extraction module and depthwise separable convolution to improve the CIFAR10-quick model;Then,the characteristics of the diseased tea leaves extracted by the improved model are input into the support vector machine(SVM)to realize diseased tea leaves identification.The multiscale feature extraction module uses convolution kernels of different sizes in parallel to fully extract the characteristics of the diseased tea leaves.Then,the multiscale feature extraction module uses depthwise separable convolution to reduce the amount of model parameters.The experimental results show that the improved model used in the proposed algorithm has the characteristics of fewer parameters,high identification accuracy and fast identification speed.The algorithm has an average identification accuracy of 0.9625 for healthy tea leaves and tea leaves infected with tea bud blight,tea leaf blight and tea red scab respectively,which is higher than the decision tree(DT)and k-nearest neighbor algorithm(KNN)and other machine learning algorithms.(3)A tea leaf’s disease identification algorithm based on Attention U-Net and WCNN is proposed.The algorithm first uses the semantic segmentation model Attention U-Net to segment diseased tea leaves;Then combine wavelet transform and convolutional neural network to build a WCNN model to identify the segmented tea leaf images.The use of Attention U-Net to segment diseased leaves can effectively eliminate the interference of complex backgrounds.Using wavelet transform to perform multi-resolution decomposition of the segmented tea leaf image can highlight the detailed information of the leaf while retaining the approximate contour of the leaf.The feature information output by wavelet decomposition is input to the convolutional neural network to extract the abstract semantic features of the diseased tea image with higher discrimination.In addition,the WCNN model used in this thesis uses the SE block after the convolutional layer to highlight abstract features useful for classification tasks and improve the identification effect.The experimental results show that this algorithm has high identification accuracy for five kinds of tea leaf images in natural scenes.