Research On Corn Disease Identification Based On Transfer Learning

Maize is one of the three major food crops in our country.Its production is not only related to the safety and satiety of the people,but also closely restricts the development of the whole national economy.The image characteristics of different maize diseases are closely similar,especially in the stage of maize leaf disease,it is difficult to diagnose and discriminate by naked eye.At present,in-depth learning research has made some achievements in the application of intelligent detection,identification and diagnosis of crop diseases in China,which are relatively mature and widely used.However,the research on intelligent detection and identification of diseases and diagnostic techniques for some crops such as rice and maize is relatively lacking.Through a new method based on convolution neural network and in-depth learning,a new intelligent detection and recognition system for maize diseases was designed and developed,which can run on mobile smartphone.The system can automatically identify and detect several common harmful diseases in Maize field,including maize big spot disease,maize gray spot disease,maize rust disease and healthy leaves of maize.The system can realize realtime,efficient and accurate identification of disease types,and help farmers better manage maize production process.The main contents of this paper are as follows:(1)Explore the basic structure of convolution neural network,and use maize leaf disease dataset to preliminary test the recognition effect of different models,and compare the traditional neural network Alex Net,Goog Le Net,Res Net-50 and lightweight neural network Mobile Net V2,respectively.The comparison results show that the lightweight neural network Mobile Net V2 has excellent recognition effect and speed,and the parameters are smaller than the traditional neural network Alex Net,Goog Le Net,Res Net-50,which is more suitable for the construction of mobile corn disease identification platform.(2)On the basis of convolution neural network,a model improvement method based on migration learning is presented.By using this migration learning method,the recognition accuracy of the model is further improved without increasing the operation time.The accuracy,accuracy,recall and specificity of the test set are combined with the overall quality performance of the confusion matrix model as key evaluation indicators of model performance,which provides an effective reference for model design and parameter selection.The results show that the fine-tuned migration learning method has better recognition accuracy.The recognition accuracy of Mobile Net V2 network after fine-tuned migration learning is 99.25%,and the single batch time is shorter than that of traditional Mobile Net V2 network by nearly 5S.The results showed that the model trained by fine-tuned migration method was more suitable for maize disease identification,and provided a new idea for building mobile disease identification platform.(3)Based on the fine-tuned migration method,a moving-end maize disease identification system was designed and implemented.Users can quickly get the information and reliability of the disease by taking photos to upload images and calling the camera to get images in real time.By testing the maize disease image in the local field,the accuracy of image recognition for photo upload is 84%,and the accuracy of image recognition for real-time acquisition is 74%.The identification system can identify several common types of maize plant with large spot disease,maize gray spot disease,maize rust disease and maize healthy leaf plants under test conditions.It provides technical support for researchers and provides important support for intelligently identifying maize diseases in China.