Research And Application Of Crop Disease Recognition Based On Convolution Neural Network

The Food and Agriculture Organization of the United Nations(FAO)estimates that plant diseases cost the world economy $220 billion in 2019.Agriculture is the foundation of our national economy and crop disease control is an indispensable link in agricultural production.Convolutional neural networks have made great contributions to agricultural development and are widely used in crop disease recognition.The recognition method based on convolutional neural network has a low error rate,but it still has some problems,such as large number of parameters,large amount of calculation and difficulty in recognizing diseased leaves under complex background.At the same time,the existing lightweight neural networks have low recognition accuracy and are not suitable for the recognition of disease images in natural scenes.In view of this,this study proposed a lightweight neural network with stronger applicability to shooting disease images in natural scenes.Specific research contents are as follows:(1)Crop disease recognition based on LAB two-branch classification model.After the convolution operation of the first three layers of the DCNN network,a multi-scale ECA module is added.Each basic block consists of multiple parallel expansion convolution with diverging expansion rates to receive receptive fields of different scales to save model parameters and calculation costs.By constructing the enhanced channel attention module ECA,the features from different channels are aggregated,so that the feature information of different locations can be extracted better without loss of image resolution.Experiments on Plant Village dataset show that the LAB two-branch classification model has a lower recognition error rate.(2)Crop disease recognition based on natural scenes.The background of crop disease images taken in actual crop planting is often complex and the accuracy of network recognition may be decreased due to non-standard image shooting.In order to make the LAB double-branch classification model more suitable for actual shooting images,double branches are added again on the basis of the network and more multi-scale ECA modules are introduced into the branches.The total sample size of crop disease images collected from the new data set was small,so image enhancement was used to reduce the overfitting speed of the model.The experimental results showed that the model recognition ability was improved after image enhancement,and the adjusted model was more suitable for cro p disease recognition in natural scenes.(3)The crop disease recognition system is designed.The modules of disease image uploading,recognition and knowledge searching were designed in the framework of wechat mini program,and the model was deployed by the Flask framework as a web service.