Research On Image Recognition Of Pests Based On Capsule Network

China is a large agricultural and forestry country,and there will be a large amount of grain and fruit harvest every year.However,plant diseases and insect pests are often encountered when planting agricultural and forestry crops,which is a very headache for agricultural and forestry practitioners.Once the problem of plant diseases and insect pests is serious,it will greatly affect the harvest of agricultural and forestry practitioners for half a year or a year.Therefore,pest control is particularly important to avoid a large area of pests and diseases on agricultural and forestry crops.With the rapid development of computer vision,many methods have been applied to pest image recognition.For example,the relatively mature Convolution Neural Network（CNN）technology developed in recent years and various Deep Convolution Network models based on Convolution Neural Network are widely used in various classification tasks.Convolutional Neural Network has no translation deformation and can extract the picture feature information well,but it also has corresponding disadvantages,such as insensitive to position information,requiring a large number of sample data during training,long training time and high requirements for equipment.Therefore,this paper studies the effect of Capsule Net on pest classification.Because of its own characteristics,Capsule Net has great advantages in dealing with small data sets.Whether using Convolution Neural Network model or Capsule Net model to classify and recognize images,there must be a training data.At the beginning of this experiment,there are only dozens of picture data of five pests.Even if the capsule network has the advantage of processing small data sets,there are still few dozens of image data as training data.To solve this problem,this paper uses crawler technology and image expansion method to expand a small amount of image data.However,there will be a problem of "what you climb is not what you need".Therefore,this paper uses alexnet to classify the crawled pictures,so as to save a lot of manpower and other costs.In this paper,comparative experiments are used to verify the feasibility of image augmentation method and Alex Net model for binary classification and labeling of crawling pests.Firstly,the original picture data is used as the training data to train five Alex Net models,and the new data set obtained by expanding the original picture data is used as the training data to train five Alex Net models again.The experimental results of the five Alex Net models trained twice are compared,and the five Alex Net models trained twice are used to classify the crawled pest pictures and compare the accuracy of classification prediction,So as to verify the feasibility of image widening method.Then,take 200 pictures of 5 kinds of pests crawled on Baidu and Sogou as training data to train Goog Le Net model,Res Net18 model and Capsule Net.Taking the pictures after the second classification of the crawled pictures using Alex Net model as training data,train Goog Le Net,Res Net18 and Capsule Net again,and compare the experimental results of the two models,So as to verify the feasibility of Alex Net model for binary classification and labeling of crawling pest pictures.Finally,the constructed training data are augmented,and trained by Goog Le Net,Res Net18 and Capsule Net.Compared with the experimental results,Capsule Net has the highest classification accuracy.Finally,the first layer convolution layer network structure of Capsule Net is improved,using the improved concept structure + 2 5 × The experimental results show that the classification accuracy of the improved Capsule Net is 4.7% higher than that of Capsule Net.