Research On Classification Method Of Vector Mosquitoes Based On Deep Learning

Mosquitoes are ubiquitous in human life,and despite their small size,they are one of the deadliest animals in the world.Mosquitoes are the vector of a variety of diseases,more than3600 mosquito species in the world,of which Aedes mosquitoes,Anopheles mosquitoes,Culex mosquitoes are the most dangerous mosquito species,Aedes mosquitoes can transmit dengue fever,yellow fever,etc.,Anopheles mosquitoes can transmit malaria,Culex mosquitoes can transmit Japanese encephalitis.Mosquito-borne diseases are occurring every year around the world,causing huge losses to human life,health and property,and in areas where mosquitoborne diseases are outbreaks,experts are required to capture a large number of mosquito samples and use traditional artificial microscopy identification methods,which are timeconsuming,labor-intensive and costly.In view of the low efficiency of manual identification of mosquitoes,this paper proposes a classification method of vector mosquitoes based on deep learning,and trains mosquito images by establishing a deep learning model,so as to achieve the effect of automatic classification of mosquito images and reduce the cost of manual identification of mosquitoes.This paper mainly studies the following two points:(1)Based on convolutional neural network,using transfer learning,fine-tuning Res Net18,Dense Net121 and Mobile Net V2 three Image Net pre-training models,using K-fold crossverification under 900 small mosquito datasets,the three mosquitoes of Aedes aegypti,Aedes albopictus and Culex mosquitoes were classified and the model performance was evaluated.The experiment set three initial learning rates were 0.01,0.001,and 0.0001,and three initial learning rates were trained on the Mobile Net V2 model to screen the appropriate initial learning rate,and the initial learning rate was set to 0.001 best.The average peak accuracy of the experimentally trained Res Net18,Dense Net121 and Mobile Net V2 models reached 95%,97%and 97%,respectively,and finally,344 mosquito images were predicted by using the retrained model under 900 mosquito datasets,among which the lightweight model Mobile Net V2 reached the highest accuracy of 0.95(Precision),recall(Recall),and F1 score.Combining the final prediction accuracy of the three models,it is concluded that the lightweight model Mobile Net V2 performs better under a small number of data sets.The experiment changed the previous model fine-tuning method,and by setting the learning rate of the model classification layer to 10 times the learning rate of the previous layer,compared with the previous experiment,the prediction accuracy of Aedes albopictus was improved by 5%～6%,which solved the problem of training convergence of a small number of data samples and further expanded the applicable environment for vector mosquito identification.(2)Based on the Swin Transformer model,transfer learning is adopted,and the hyperparameters of the model are the same as those set when fine-tuning the convolutional neural network,and the K-fold cross-validation method is also used to verify the feasibility of the hyperparameters selected for the model on the 900 mosquito dataset,which is used to evaluate the performance of the Swin Transformer model under a small number of datasets.During the cross-validation,the average peak accuracy of the trained model reached 98%,which was about 1% higher than the average peak accuracy of Dense Net121 and Mobile Net V2 in the previous chapter experiment,indicating that the Swin Transformer model performed better than the former convolutional neural network model with a small amount of data sets.Then the Swin Transformer model was retrained on the entire 900 mosquito dataset,and 344 mosquito images were predicted by the final trained model,and the accuracy of the three predicted indicators,recall and F1 score were all 0.97,which was higher than the 0.95 of Mobile Net V2,and the comparison showed that the Swin Transformer model was less affected by a small number of datasets.Finally,the experiment calculates that the overall prediction accuracy of the model for the three types of mosquitoes is 96.80%,which is 2% higher than that of Mobile Net V2 of 94.77%,indicating that Swin Transformer has a better classification effect on mosquitoes than the convolutional neural network model,and Swin Transformer can replace the convolutional neural network model and be used in the classification of mosquitoes.This paper aims to solve the time-consuming and laborious problem of traditional mosquito identification methods,and to solve this problem,the vector mosquito classification method based on convolutional neural network and the vector mosquito classification method based on Transformer are discussed,and the efficiency of mosquito classification is improved to the greatest extent through the comparison of the two methods,and experiments show that the performance of Transformer in mosquito classification is better than that of convolutional neural network.