Crop Disease Recognition Method Based On Deep Learning

Diseases of crops will greatly reduce the output of crops and affect the quality of crops.Therefore,it is necessary to identify the diseases and take timely control measures according to the types of diseases to achieve high quality and high yield of crops.In recent years,the maturity of deep learning technology makes it possible to use deep learning algorithm to identify diseases intelligently.In this dissertation,the convolution neural network is used to train the data set to recognize the disease.The main research contents include::(1)Using crop disease data set from AI Challenger competition,sharpening,filtering and other operations to enhance image data features,reduce the adverse impact of image noise error on the experimental results,and label the data set images.(2)Considering that the depth of multi-layer network in depth learning is easy to increase the training complexity and the gradient disappears,the ResNet convolution neural network model is used to train the data set of crop diseases.The residual learning unit in ResNet network can simplify the learning process of convolution layer,only learn the difference between input and output,deepen the network layer and accelerate the model training.During the training,BN layer is added in front of each layer of residual learning unit to avoid the gradient disappearing,and small batch gradient descent method is used to optimize the network parameters.At the same time,the recognition accuracy of ResNet in the convolution layer of 50,101 and 152 is compared.(3)In view of the phenomenon of network performance degradation and over fitting in training ResNet model,the dissertation adopts the inception_v3 neural network with sparse structure and high computing performance,and combines the factors of a large number of data set images,and combines the inception_v3 network with migration learning strategy to train and identify disease images.The disease feature of the image is extracted from the bottleneck layer through the inception_v3 network.The disease feature vector with strong expression ability is saved in the bottleneck layer file.The bottleneck layer file is used as the input of the subsequent network layer to fine tune the training parameters.The gradient descent method is used to update the weight of the network to effectively alleviate the over fitting.The disease category recognition is realized through softmax.The experimental results show that the recognition accuracy of the model is 93.90%,which shows that the accuracy of the model for disease recognition is effectively improved.(4)Building the practical application platform of algorithm.At the back-end of the platform is the training results of the model of the combination of the inception_v3 network and the migration learning,which is used to test the categories of crop diseases.The server is built on the Django framework,the front-end is mainly the static test and result display page,and the back-end is the training results of the inception_v3 model.The front-end and the back end are integrated,and the pictures to be tested are loaded on the test page of the server,the back-end is judged,and the results of the back-end test are presented on the recognition results page.