Recognition Of SEM Images Of Atmospheric Particulate Matter Based On Convolutional Neural Networks

Under the background of global concern about air pollution prevention and control,PM2.5 accounts for a large proportion of air pollution.It not only directly or indirectly affects the global climate,but also poses a great threat to human health.Therefore,the study of PM2.5 has become an important topic of air pollution prevention and control.This thesis mainly classifies and identifies the scanning electron microscope image of particulate matter based on convolution neural network,and provides a theoretical basis for the source apportionment of atmospheric particulate matter by studying the morphological characteristics of PM2.5 single particulate matter.Convolutional neural network is one of the commonly used frameworks for in-depth learning.By using multi-convolution kernels to convolute image data,important features of images can be obtained.Therefore,convolutional neural network is widely used in image.Through a large number of literature reports and current applications,convolutional neural networks can be used to identify particles with different morphological characteristics in PM2.5.Seven sampling sites in Tsingtao were selected during the study.A large number of scanning electron microscope images of PM2.5 were collected intermittently from 2016 to 2018.Four types of particles with obvious shape characteristics were selected and classified into spherical,flocculent,mineral and fibrous particles.Then the morphological characteristics of the particles were observed,and the chemical composition of the particles was analyzed according to EDX energy spectrum analysis report.Through the convolution neural network,3114 samples were trained and 692 samples were used as test sets.The recognition rate of training samples was 99.65%,and the overall recognition rate of test samples was 98.84%.The final results show that the network structure has a significant effect on the recognition rate of the experimental samples,and achieves the ideal classification effect.The feasibility of convolution neural network combined with SEM image for particulate matter recognition in PM2.5 is preliminarily verified.