As one of the most important food sources for human beings,the growth process of crops is extremely susceptible to external factors such as pests and diseases.Effective and timely pest control measures can help restore a large amount of food losses,and play a vital role in the agricultural economy and sustainable agricultural development.Accurate and rapid identification of pests is a key step for effective pest control.Traditional pest identification methods mainly rely on human experts to perform classification work based on their own experience or access to information,and accuracy and timeliness are difficult to meet the current application needs of intelligent agriculture.With the rapid development of science and technology,the application of artificial intelligence technology in the agricultural field has gradually become widespread.Pest identification based on deep learning methods has shown strong performance advantages and provided new ideas for agricultural pest control.In order to solve the problems of low efficiency and low accuracy of traditional agricultural pest identification,this paper uses deep learning methods to construct an agricultural pest image data set HAULP31 and train it to obtain an optimal pest automatic identification model.And preliminarily deploy remote data collection equipment,try to use this model to classify and recognize pest images in the real wild farmland environment.The main work is as follows:(1)Based on the research on agricultural pests in the laboratory,we photographed and collected the adult images of 31 agricultural pests,annotated the images with the CVAT annotation tool,and carried out the cutting and data enhancement of the individual pest images.Finally,a set of 44881 images was established.The HAULP31 dataset of agricultural pest images.The data set in this paper has high credibility,and has a certain degree of versatility and generalization.It can be used for a variety of insect recognition tasks and migration learning,providing a data basis for the follow-up research of the research group.(2)Based on the Tensor Flow learning framework and HAULP31 data set,Mobile Net,Res Net50V2,Dense Net121 and Efficient Net B0 are selected as experimental network models to study the classification and recognition technology of pest images.After comparing and analyzing the experimental results of the four networks,Mobile Net is superior to the other three networks in terms of weighted accuracy,training speed and other data.It achieves a weighted accuracy of 98.89% on the HAULP31 data set,which proves deep learning.The effectiveness of the model in the field of pest identification.(3)In view of the excellent performance of Mobile Net on the HAULP31 data set,this paper adjusts and optimizes the hyperparameters based on the three gradient descent optimization algorithms of RMSprop,SGD,and Adam to continue to explore the modeling advantages of Mobile Net on the HAULP31 data set.The performance of Mobile Net optimized by different hyperparameters has been improved in different ranges,and its highest weighted accuracy rate reached 99.76%.At the cost of training time,the classification and recognition capabilities of the model were further improved.(4)In order to further test the practical application capabilities of the model,remote insect monitoring equipment was initially deployed in the real field farmland environment to conduct experiments,and data collection was achieved by automatically trapping insects and shooting their images and transmitting them to a remote server.Using the Mobile-LP31 model to classify and recognize pest images in the wild environment,the weighted accuracy is 75.11%.The results show that the model in this paper has a certain practical application effect,and the HAULP31 data set has a certain modeling value for pest control in agricultural production. |