The Research Of Tea Buds Detection And Leaf Diseases Image Recognition Based On Deep Learning

Convolutional neural network(CNN)is a widely used model of deep learning.It can effectively reduce network complexity by its own weight sharing,local connection and pooling operation.It can not only give the model a certain degree of invariance in translation,distortion,scaling and other operations,but also show strong robustness and fault tolerance.In addition,convolutional neural network can simultaneously extract features and recognize patterns,which can avoid complicated explicit feature extraction process,so it is easier to train and optimize the network.In recent years,CNN has played an important role in face recognition,agricultural disease image recognition and other fields.Target detection can be understood as an image segmentation based on geometric and statistical features of the target,which can be divided into target classification and target location.The task of target classification is to determine the target category in the input image;the task of target location is to determine the location of the target in the input image.In recent years,especially in complex scenes,it is very important to recognize and locate real-time targets automatically when multiple targets are detected.YOLO algorithm uses convolutional neural network to integrate feature extraction network,location frame prediction and category prediction into a framework and achieves end-to-end training.It can automatically learn tasks,realize multi-layer non-linear transformation and high-level abstract description of images.In order to improve the accuracy of tea bud image detection and leaf disease image recognition,the related research work of target detection and convolution neural network is carried out.The main research contents are as follows:(1)A method for tea bud detection based on YOLOv3 algorithm named Darknet_tea is proposed.It was first applied to the detection of tea bud image under complex background,and improves the YOLOv3 network architecture from the aspect of multi-scale detection.The original YOLOv3 algorithm performs the regression of detection frames and target categories in three different feature response maps,and a single grid predicts nine target frames of different sizes,resulting in a large amount of computation of the model.Considering that the tea bud area taken in the field occupies a larger position in the image,and the tea bud and the old leaf have different shapes and colors.So based on the super green feature of the extracted image and the image segmentation using the OSTU algorithm,the Darknet_tea algorithm proposed in this paper only predicts regression on the scales of 13×13 and 26×26,which not only improves the accuracy of tea bud detection,but also reduces the computational load of the network.(2)A method for the image generation of tea leaves disease based on Deep Convolution Generative Adversarial Networks(DCGAN)was proposed.In the collection of experimental data,due to the different degree and time of occurrence of each disease,the number of some diseases was less.Therefore,DCGNN was used for data enhancement to expand the image data of leaf disease and balance the number of each disease species,which lays a foundation for the recognition of tea leaf diseases.(3)A method for leaf disease images recognition of tea based on CNN model was proposed.To prevent the appearance of over-fitting,Dropout and Local Contrast Normalization(LRN)are added.In order to increase the sparsity of the network and eliminate the gradient dispersion,the Rectified Linear Unit(ReLU)is used as the activation function in the model.Therefore,the convergence of the network is accelerated and the structure of the network is improved.