Research On The Convolutional Neural Network Algorithm Of Disease And Insect Pest Identification In Ningxia

Crop pest control is a key link in agricultural production.Accurate image recognition of pest has the characteristics of high efficiency,low cost and convenient operation.It is the main technology of pest control in recent years.Because of its unique geographical advantages,Ningxia has long winter and short summer.It has unique environmental conditions,such as large daily temperature difference,long sunshine time and strong solar radiation.These are suitable for the growth of grape,corn,potato,medlar and other crops.These crop have effectively boosted the economy of the Ningxia Region.The above crops are easy to be attacked by many kinds of diseases and insect pests.Because of their abundant starch and sweet fruit,which have serious influence on fruit yield and quality.Based on the image recognition method of convolutional neural network(CNN),this paper improves the Yolo algorithm and proposes a novel Yolo+K-Means++recognition algorithm to improve the recognition accuracy and make the recognition result more reliable.Experimental results showed that this algorithm identified 15 pests and 11 diseases effectively,and has high recognition efficiency and accuracy.This algorithm lays a solid foundation for precise control of crop diseases and pests.The works which did by this paper are divided into the following points:(1)Collecting and making the image data set of insect pest and leaf disease.According to the research requirement,19 kinds of pest pictures were collected,including 3634 pictures,and 20 kinds of leaf pictures of different diseases were obtained,including 37200 pictures.After image preprocessing and batch normalization,15 insect pests and 11 disease images were selected as the research objects.In order to improve the recognition accuracy,enhancing the usability of the images,the data set was expanded by the methods of spatial transformation,sample enhancement and normalization.The images of 15 insect pests were expanded to 25304,and the images of 11 leaf diseases were expanded to 18300.The expanded dataset contains 43,604 images.(2)Setting up the development environment,selected the suitable recognition algorithm of diseases and insects to do the experiment.By comparing the principles of different algorithms and the experimental results of predecessors,the network models based on MobileNet V2 and Yolo algorithms are selected for training.By comparing the recognition accuracy,time and final effect of the two models,Yolo model is more suitable for the recognition of this paper,so we choose Yolo as the basic network,and improve the model on this basis.(3)The method of combining K-Means++with Yolo algorithm is proposed to solve the problem of insufficient result of Yolo algorithm.The improved algorithm uses anchor boxes to detect the target.The length and width of anchor boxes are obtained by K-Means++clustering algorithm before detection,the Yolo algorithm can get the most likely target region from the labeled data by clustering statistics,and enhance the recognition accuracy.The results show that the new network improved the recognition accuracy of insect pests and leaf diseases,and solve the problem of complex scene and multi-object image recognition.(4)Training the data set of disease and insect pest on the innovative Yolo+K-Means++recognition algorithm,through many times of training,and comparing the experiment with other algorithms,prove the superiority of innovation network.This paper designed and implemented a GUI page to facilitate future application of insect pest identification to the ground.After research and experiment,Yolo+K-Means++recognition algorithm has a detection accuracy of 98.618%for 15 kinds of insect pests and 99.218%for 11 kinds of leaf diseases.Because of the difference of data between the two kinds of images,the total recognition accuracy of the algorithm is 98.87%.The results show that the new algorithm is superior than other algorithms,and the precision of the Yolo algorithm is improved by 11 percentage points.The design of GUI page also lays a foundation for the realization of mobile application of pest identification in the future.