Research On Fine-grained Visual Classification Of Plant Disease Based On Transfer Learning

China is a traditional agriculture country,in recent years,the total value of china’s agriculture import and export has been growing continually,and plant disease recognition is an important issue in the agriculture industry.The method of artificial recognition and traditional computer vision has limits in practical applications.So it is necessary to research plant disease recognition algorithms based on CNN(Convolutional Neuron Network).The method of artificial recognition and traditional computer vision is hard to adapt practical application,an CNN model can fit distinct tasks through parameter optimal,so it is necessary to study plant disease and insect pest identification algorithms based on CNN.Based on the existing plant disease image data,this paper first use the traditional CNN model to evaluate the effectiveness of transfer learning for plant disease FGVC(Fine-Grained Visual Classification)task,then use CNN based on discriminative region attention mechanism to evaluate the dataset,then fusion discriminative region attention mechanism model and channel attention mechanism,this paper use this fusion attention mechanism to evaluate the dataset,then construct plant disease recognize system based on the fusion attention model.Our main work and result as follows:(1)Based on existing plant disease image data and the basic theory of CNN,use CNN models such as Alex Net,mobile Net and Res Net to evaluate the plant disease dataset,in the training process this paper use the data augment strategy to expand the dataset,and the transfer learning strategy is used to solve the issue of few data,and the effectiveness of the strategy is verified by our comparative experiments.(2)The CNN based on discriminative region attention mechanism is proposed for the plant disease FGCV,it uses the fine-grained feature of the image to improve classification accuracy by extracting the discriminative region of the target.This paper verify the effectiveness of discriminative region attention mechanism on plant disease FGVC,then combine the discriminative region attention mechanism and the channel attention mechanism to cause attention relation between channels to the deep feature of CNN,then verify the effectiveness of this fusion attention model,the accuracy is significantly been improved by 1.9%.(3)This paper construct a Web-based plant disease recognize system,then deployed the system to a cloud server.It makes the CNN model that has high environmental requirements can compatible with various terminals,in the practical application the user can achieve various functions by accessing the web: for the framer,they can use the algorithm that deploys on the server to recognize the plant disease by taker or upload target photo to the server,consider the needs of the agriculturist,the system can detection and statistic the disease parts location of plant disease dataset by using CNN model based on fusion attention mechanism.