Application Of Deep Convolutional Neural Network In The Identification Of Maize Leaf Diseases

Maize is widely planted and has a high total output in China.It is an important food crop and feed crop for the national economy.However,due to changes in farming and cultivation systems,variations in pathogenic strains,and imperfect plant health measures,maize leaf diseases have occurred and the degree of harm has increased,and the types have also increased.This problem has always restricted the increase of maize production and brought great economic losses to farmers.Traditional identification methods often rely on personal experience and visual observation.This method has the characteristics of slow identification speed,strong subjectivity,high misjudgment rate and poor real-time performance,which can not meet the needs of agricultural development in China.Some methods first preprocess the image,and then use machine learning methods such as Support Vector Machine(SVM)and Probabilistic Neural Network(PNN)to classified the disease.The image processing process of these methods is complicated,and the identification accuracy needs to be improved.Therefore,timely and accurate diagnosis of maize leaf diseases can not only reduce unnecessary financial expenditures and resource consumption,but also obtain a considerable corn yield in a changing environment.In this paper,the following problems are mainly solved for the identification of maize leaf diseases based on deep convolutional neural networks:(1)Research on the effects of structural changes in deep Convolutional Neural Network(CNN)models on the accuracy of maize leaf disease identification,such as changing different pooling combinations,adding Re LU functions,and dropout operations(exploring the effects of different dropout thresholds on model performance),selecting the appropriate classifiers classify the results,etc.,and discuss the improvement of the Cifar10 and Goog Le Net models.The improved top-1 average recognition accuracy of the Goog Le Net model is 98.9%,and the average accuracy of the Cifar10 model is 98.8%.The improved model has better identification accuracy and lower model loss than the original.(2)Research on the number of images in the image dataset is constant,the performance of the model is adjusted after adjusting the hyperparameters by using some hyperparameter adjustment techniques.The hyperparameters include: “Train batch_size”,“Test batch_size”,“lr_policy”,“base_lr” and "weight_decay".This research is based on CNN models: Alex Net,Goog Le Net,and Res Net18.When using the above three structures to train and test the maize leaf image dataset,the test accuracy after the model hyperparameter adjustment was increased by 6.075%,6.05% and 6.4%,respectively;the generalization test accuracy was increased by 12.85%,7.025% and 11.95%.(3)When the number of test images is constant,uses data augmentation methods to expand the original 1129 training images by 2 times(2258),4 times(4516)and 8 times(9032).Research on the impact of different numbers of training images on the performance of different CNN models.For the Alex Net,Goog Le Net,and Res Net18 models,as the number of images increases,the average test identification accuracy increases from 82.6% to 93%,from 56.3% to 91.8%,and from 85.6% to 92%.The phenomenon of over-fitting gradually disappears and the performance of the model is more stable.(4)Research on the performance of different CNN models for different image datasets.The experimental results show that the relationship between network structure(network complexity,network depth,number of parameters included in the network,etc.)and model performance is weak.