Study On Recognition Of Soybean Disease Based On Convolution Neural Network

Crop diseases can have a serious impact on crop yield across countries all over the world.How to accurately identify different kinds of diseases and scientifically diagnose diseases are important research directions of intelligent agriculture.This thesis will select virus disease,powdery mildew,downy mildew,brown patch and gray spot disease,the five most common diseases of soybean diseases,as research objects and study the automatic identification of the disease.The identification of crop diseases is mainly based on the characteristics of artificial extraction of colors,shapes and spectrum at present.However,these characteristics are easily influenced by natural environment and human factor.In order to solve the problems mentioned above,this thesis will adopt the method with the characteristic of convolution neural network and studies the automatic extraction of disease characteristics.With the strategy of improving and compressing models,this study achieves a better effect in the smaller model size,which provides a theoretical support for the application of automatic identification of soybean diseases in low-end mobile platform.The main research of this thesis includes the following aspects:(1)Constructing the training data,set of soybean disease model.By using sample expansion method like enhancement,flip and adding noise,this study expands the five disease data set to 3430,which compensates for the shortcomings of small training samples.A pre-training convolution neural network model has been put forward in this thesis to solve the problem of small sample size.By adopting the vector quantization of ideas and building characteristic expression of the dictionary,this thesis regards the clustering results as Convl weights,constructs various convolutional neural network models,analyzes the distribution characteristics of weights and discusses the influence of the optimal K value,Patch size,activation function and different initialization modes on the network.The research model combines with the strategy of Dropout data enhancement to solve the problem of over-fitting and reduce the influence of network depth on the effect of K-means pre-training.This model solves the problem of the defect in the random initialization weight and verifies the validity of the K-means pre-training method with the small samples of soybean disease.As a result,the accuracy of this model achieves 96.7%.(2)According to the unsupervised pre-training neural network,this thesis proposes a simplified network architecture model based on batch standardization to deal with the problem of the large model,slow convergence and low stability.By adding BN layer and adjusting its numerical distribution before activating function,it solves the problem of slow speed of convergence,neuron necrosis and so on.According to the characteristics of convolution neural network architecture,this thesis sets different convolution kernel number and size and different FC nodes to reduce redundancy and sets threshold value to filter the weight of small actions with the strategy of weight clipping and architecture compression.This model can be used to validate in the recognition model of soybean diseases.And this model finally compresses the parameter memory to 9.6M,which makes the accuracy quickly converges to 94.6%.This model not only guarantees the precision of the model but also greatly reduces the model sized and the amount of calculation.The advantages of this model will help researchers to solve the problems that the current model memory and compulation are too large to be applied to the low-end computing platforms such as mobile terminals and agricultural robots.