The Research Of Wheat Leaf Disease Images Identification Based On Convolution Neural Network

Convolutional neural network(CNN)is a variant Multi-Layer Perception(MLP)inspired by the Hubel-Wiesel biological visual system.The implicit description of image feature can be adaptively constructed by multi-layer nonlinear mapping,which doesn't need to change the topology of the image.The layer-by-layer analysis of features can be realized through local receptive field,weight sharing,subsampling.At the same time,the number of parameter can be reduced and the extracted image features are robust to the rotation,translation,scaling and other transformations.What's more,feature extraction and pattern recognition are carried out at the same time,avoiding explicit feature extraction and data reconstruction of surface learning.In recent years,the remarkable achievements has been achieved by CNN in the fields of face recognition,fingerprint recognition and others.Aiming at improving the accuracy of wheat leaf disease image identification,the main research of CNN are carried out in this paper.The main contents are as follows:(1)A method for identification of wheat leaf diseases based on Fine-grained differential amplification convolutional neural network(F-gDACNN)is proposed.It mainly has 2 improvements: convolutional kernel matrix and differential amplification branch.The former can effectively increasing the number of neurons and link channels and suppressing parameter expansion.At the same time,the feature maps extracted from different filters can fuse with each other,which can produce new features;The latter can amplify small differences between the real output and the expected output,so the weight updating become more sensitive to the light errors in the backpropagation pass.This structure significantly improves the model's fitting capability and reduces underfitting.(2)A method for identification of wheat leaf diseases based on multichannel convolutional neural network(MCNN)is proposed.Inspired by the human visual behavior in video saliency detection,the branches are added on the basis of the serial structure.The first and second subsampling layers are connected directly with the first fully connected layer,which realized the comprehensive utilization of the image features in different levels.In addition,the mixed modes of rectified linear units(ReLU),Dropout,local contrast normalization(LRN)and local response normalization(LCN)are introduced to prevent overfitting and gradient diffusion and make the network structure more perfect.(3)A method for identification of imbalanced wheat leaf diseases based on CNNLSVM is proposed.In view of the low accuracy of CNN in imbalanced data identification,a CNN-LSVM model is proposed by combining local support vector machine(LSVM)with CNN in this chapter.Then,considering the imbalance in the number and spatial distribution of the samples,a cost sensitive matrix is designed to assign the value for penalty factors in the optimized objective function of LSVM.It makes the model more sensitive to the misdivision caused by the imbalances of the samples and modify the convergence trend of the model,which makes the model more suitable to the classification of imbalanced datasets.This model not only preserves the advantage of automatic abstraction of abstraction features by CNN,but also integrates the improved LSVM to alleviate the problem of imbalanced classification and to make full use of the local information of samples.