Research And Application Of Recognition Algorithm Of Rice Diseases And Insect Pests Based On Convolution Neural Network

Rice diseases and insect pests account for as much as 50% of the yield loss.Therefore,it is necessary to adopt effective and economic sterilization and insecticide treatment,whose effect depends on the type,situation and time of infection.In these cases,it is necessary to identify the specific diseases and insect pests correctly and early in order to minimize the loss of yield and improve the efficacy and efficiency of treatment.In recent years,many methods of crop disease automatic recognition based on image analysis have been proposed.Among these methods,the application of deep convolution neural network(CNN)has been proved to be very successful in different visual classification tasks.In this paper,based on the deep learning technology,this paper studies the identification of rice diseases and insect pests.Firstly,a data set of rice diseases and insect pests is sorted out by crawling the images of rice diseases and insect pests on the Internet and collecting the pictures on the spot.The data set contains 15 kinds of diseases and 22 kinds of pests.It is divided into training set,validation set and test set,which are used to identify the training,verification and test of the network respectively.In order to make up for the large difference in the number of different types of pictures and improve the generalization ability of recognition network,this paper uses conventional image processing,mixup algorithm and generative adversarial networks to enhance the data set.Then,this paper proposes a FRNet network based on the depth residual neural network for rice and pest identification.First of all,regularize the feature maps of the last convolution layer of the residual network,and then add the weight to the loss function,so that the learned characteristics of each feature map are as different as possible.Experiments show that the recognition accuracy of FRNet network is improved compared with Res Net network,and the effectiveness of the improvement is verified.Based on the feature that SENet network can automatically assign different weights to each feature map,this paper introduces this feature into the weight assignment of the last output feature map of FRNet,which not only increases the difference between the feature maps,but also automatically assigns the important feature map to the large weight and the secondary feature map to the small weight.Experiments show that the improved FRNet network is 0.13% and 0.29% higher than the original one in disease and pest verification sets,respectively,which proves that the FRNet network and its optimization for FRNet proposed in this paper are effective.Then we propose a model pruning method based on L1 norm of convolution kernel.The improved vgg-16 model is tested with different pruning proportion,and then the model after pruning is retrained,so that the recognition model with lower pruning proportion can be restored to the same recognition accuracy as before pruning.The model after pruning has faster recognition rate than the model without pruning,and takes less hardware resources,and achieves the goal of model lightweight.Finally,a rice disease and insect identification system is designed and implemented based on the pruning and training rice disease and insect identification network.The system is composed of the server and the client.The server is deployed on the cloud server.The client has three different carriers: Web site,android app and wechat applet.Each carrier can upload pictures and feedback the recognition results.The feedback results include the name,picture and similarity percentage of the disease and pest,and can further check the details of the disease and pest distribution,symptoms,occurrence rules and control measures,so as to deepen the user’s understanding of the disease and pest and facilitate the farmers to take corresponding treatment measures in time.So it has important practical significance to improve the yield and quality of rice.