Research On Mosquito Swarm Count Technique Based On Deep Learning Algorithm

Monitoring the density of mosquitoes and other insects that carry and transmit pathogens is crucial for predicting the spread of viruses and developing appropriate measures to prevent their transmission.Mosquito counting is an essential part of this work,but existing methods rely largely on manual labor,which is inefficient and puts personnel at risk of being bitten.Therefore,developing an efficient and contactless counting method is of great practical significance.In this thesis,we investigate mosquito counting techniques in digital images from the perspective of deep learning.The main contributions of this thesis are as follows:(1)We introduce the current research status of mosquito counting and analyze several common target counting methods.In cases where some counting methods perform poorly,we develop into the counting principles of density map regression methods based on deep learning and introduce the relevant knowledge of deep learning.Besides,we collect and label a mosquito dataset,which includes 391 mosquito group images with a total of 15,466 mosquitoes labeled.(2)We propose a multi-scale mosquito counting method that incorporates an attention mechanism.The proposed method consists of front-end and back-end networks.The front-end feature extraction network is constructed from a modified VGG-16 network and is mainly used to extract low-level features of mosquitoes.The back-end multi-scale regression network consists of three parallel convolutional layers with different kernel sizes,which are used to extract high-level features of mosquitoes.In addition,a channel attention module is introduced in the third pathway to enhance the feature extraction of small mosquitoes.This model maps mosquito swarm images to their density maps to estimate the number of mosquitoes.In the experimental results,the mean absolute error(MAE)of the count was 1.810,and the root mean square error(RMSE)was 3.467,indicating that the proposed method performs well in accurately counting mosquitoes in the dataset.However,the generated images lack quality constraints,some of the predicted density maps contain noise.The image’s peak signal-to-noise ratio(PSNR)of 35.786 and a structural similarity index(SSIM)of 0.944.Indicating that there is room for improvement in both image quality and counting accuracy.(3)To address the issues with the method proposed in(2),we further propose a mosquito counting method based on generative adversarial networks(GAN)that aims to improve the quality of predicted density maps.Based on the previous model,we deploy the attention-based multi-scale regression network as the generator model and introduce a discriminator model to constrain the generator network to generate high-quality mosquito swarm density maps,thereby further improving counting accuracy.Additionally,we replace ordinary convolutional layers in the back-end multi-scale regression network with deformable convolutional layers to enhance the modeling ability of the convolutional layer for mosquito geometric deformation.Experimental results show that the proposed method achieves an MAE of 1.502,RMSE of 2.660,PSNR of37.108,and SSIM of 0.976 on the mosquito dataset,which outperforms the previous method.