Convolutional Neural Network Based Study On Classification Of Maize Leaf Diseases

Maize is an important food crop in China,which has a wide planting area and high yield,and provides great help for China’s economic development.However,due to the artificial farming system,varied pathogenic bacteria,weather and other reasons,the diseases of corn leaves are gradually aggravated,showing complex and diverse diseases.This problem has always been the primary problem affecting corn yield and causing economic losses.Therefore,the identification and classification of maize leaf diseases has always been the focus of attention.The traditional recognition method is artificial naked eye observation,but detecting leaf disease through visual observation requires an expert group and continuous crop monitoring.Therefore,it is very expensive,time-consuming and unreliable.And can not meet the requirement of real-time.Until the rapid development of machine learning,the corn images were preprocessed,and then the corn leaf diseases were classified by support vector machine and neural network.However,the process of this method is complicated,the training time is long,and there is much room for improvement in recognition accuracy.Timely and accurate recognition of corn leaf diseases can not only reduce unnecessary economic losses,but also increase the total yield of corn.Therefore,in this thesis,the convolutional neural network algorithm in deep learning is used to classify corn leaf diseases,and the research contents of this thesis are as follows:(1)By analyzing the performance of several popular network models,such as AlexNet,VGG-16 and GoogLeNet,it can be seen that the GoogLeNet model can identify corn leaf diseases more accurately than the other two models,and the effect is better.(2)In the experiment,it is found that the loss function convergence of GoogLeNet is slow in the training test.The GoogLeNet model is improved,and the batch standardization layer is added after the convolution layer and before the pooling operation,which makes the data return to normal distribution and reduces the training time.It was found in the experiment that the training set performed well when the features were activated by the first two neurons,but the deep leaf features in the test set were not distinctive enough.In order to make the features trained by the network more distinguishable,the Center loss function is added,and the Softmax loss function is matched with a certain coefficient to improve the classification accuracy.Through experiments,it is found that the optimal coefficient of the model is 0.09,the adjusted learning rate is 0.001,and the learning rate is updated by step method.At this time,the recognition accuracy of the model reaches 96.64%,which is 1.21% higher than that of the original model;(3)In order to verify the good performance of the improved model in identifying maize leaf diseases,OpenCV was used to enhance the data,simulate the leaf shape under different light and noise conditions,and generate a new test set.By comparing the experimental results,the average recognition accuracy loss of the improved model under the influence of extremely complex background is 4.12%,which is lower than that of other models.It can be shown that the improved model has good robustness.