Particle Pollution Estimation From Image Based On Convolutional Neural Network

Since particle pollution has severely affected people's health and daily activities,monitoring is of great importance.However,the traditional methods,which are mostly based on professional monitoring stations and need enough monitoring points to keep the accuracy,are faced with problems like human resource,material supplement and financial limitation.Recently,as the rapid development of image devices and deep learning technology,the particle pollution estimation based on convolutional neural network technology came into focus,which can not only improves the detection efficiency,but also reduce the human and material cost of detection.Therefore,we proposes several image-based particle pollution estimation methods based on convolutional neural networks in this thesis.The main contents are as follows:(1)A new dataset of image-based particle pollution estimation is constructed,including two sub-data sets named SHFW and BJFW.Each data includes a natural image,four-dimensional weather features(humidity,temperature,wind speed,and pressure)and corresponding PM2.5 values.(2)A neural network,based on residual neural network(ResNet)and aiming at getting finegrained estimation of image-based particle pollution,is proposed.Compared with the traditional machine learning algorithms SBFS on SHFW dataset,the root mean square error(RMSE)is decreased by 47.43%,and the R-squared goodness of fit increased by 52.63%.Further,using the transfer learning methods to optimize the image-based particle pollution estimation model,which decreases the RMSE by 2.57% and increases the R-squared goodness of fit by 2.30%.(3)A shallow ResNet,PMIE,with layer enhancement for particle pollution estimation,is proposed.Considering about different attention of image object recognition and image-based particle pollution estimation task,a shallow ResNet is designed.Besides,we proposes an inter-layer weights analysis of convolutional neural networks method,and a new method for enhancing the effect of the convolution layer.Compared with the ResNet,the proposed PMIE is more suitable for image-based particle pollution estimation task,and it decreases the RMSE by 11.86% on the SHFW dataset and increases the R-squared goodness of fit by 4.59%.And on the BJFW dataset,the RMSE is decreases by 14.38%,and the R-squared goodness of fit increases by 23.63%.(4)An image-based particle pollution estimation method,with weather features,is proposed.The four-dimensional weather features of humidity,temperature,wind speed and pressure are integrated into the image features learned from PMIE to optimize the image-based particle pollution estimation model.After merging weather features,the RMSE decreases by 0.094% on the SHFW dataset and the R-squared goodness of fit increases by 1.09%.And the similar improvement happens on BJFW dataset,in which RMSE decreases by 5.33% and the R-squared goodness of fit increases by 6.60%.In this thesis,we proposes several image-based particle pollution estimation methods based on convolutional neural networks.The methods improve the efficiency of particle pollution detection while reducing cost.