Research On Crop Pest Identification Based On MobileNet V2 Parameter Transfer Learning

Traditional pest identification methods have low accuracy and take a long time.Convolutional neural network(CNN)is widely used in the field of crop diseases and insect pests recognition because of its characteristics of automatic image feature extraction,strong generalization ability,high recognition rate and short time-consuming.Transfer learning can quickly and accurately identify crop diseases and insect pests by using the similarity of tasks and data.Ningxia has the largest output of Lycium barbarum in China,and is also suitable for grape,potato and other crops.In the process of production and planting,it will be damaged by various diseases and pests,which will lead to the decline of quality and yield,and affect the economic development of Ningxia.In this paper,the research on the identification of diseases and pests can accurately identify the types of diseases and pests,and promote the prevention and control of diseases and pests.In this paper,we use the transfer learning model to identify 11 kinds of pests,grape,potato and other 6 kinds of diseases.The experimental results show that the migration learning algorithm can effectively identify the above 17 diseases and insect pests,and lay a solid foundation for the prevention and control of crop diseases and insect pests.This paper focuses on the following three aspects:(1)Collect and expand the data of diseases and insect pests.Through field and network collection,2531 pieces of 11 species of Lycium barbarum pests including green leafhopper and inchworm and 6 kinds of disease data were collected and selected for experiments.After preprocessing the selected data,25310 pieces of pest data,13398 pieces of disease data and 38708 pieces of picture data were obtained;(2)Pest image recognition based on mobilenet V2 parameter migration model.Through the analysis of the model structure,mobilenet V2 is selected for pest identification,and the accuracy of Inconcept V1,Shufflenet V2 and Resnet-50 for pest identification is compared.Resnet-50(93.46%)and mobilenet V2 models(94.24%)with high accuracy of the verification set are selected to train the weight parameters on the Imagenet data set,then the network model obtained after training is migrated,and fine-tuning training is carried out on the enhanced target data set,The accuracy of resnet-50 and mobilenet V2 parameter migration model on the validation set of this pest data set is 98.99%and 99.27%respectively.Compared with mobilenet V2 model,the recognition accuracy of mobilenet V2 parameter migration model on this data set is improved by 4.87%;(3)The identification results of plant diseases and insect pests based on mobilenet V2 parameter migration model are displayed,and the trained model is tested on the web page.