Research On The Intelligent Classification Of Fog Grade And The Estimation Visibility Based On Deep Transfer Learning

With the development of economy,in recent years,the fog weather is more frequent,which has brought great harm to people's life.The traditional equipment for fog detection and prediction based on remote sensing technology often has many shortcomings,such as low spectral resolution,high price,large cumulative errors,and uncertainty in single-point detection,which lead to unsatisfactory result.Fog detection based on machine learning has many shortcomings,such as the incomplete feature extraction of fog image,the invariant feature space,the classification estimation criterion with insufficient approximation ability,the low posterior evaluation characteristics of fog grade and visibility classification estimation results.Those disadvantages make the generalization of fog image with irregular diversity poor.In order to solve the problem that the traditional fog measurement equipment has poor performance,the transfer learning thought and closed loop control are introduced to deep convolution neural network.Finally,an intelligent classification estimation model for gradig and estimating fog based on transfer learning has been designed.The main part in this paper is listed as follows:(1)This paper introduced the dynamic interleaved group convolutions into VGG16 network,obtaining a new classification model.This new model has low parameters,uses small saving space and has high operating speed.(2)Aiming at solving the problem of redundant information caused by convolution operation in VGG16 model with interleaved group convolution,the model introduced the Markov separability measure function based on principal component and classification / estimation decision information system to build compressed feature space.Finally,the feature set with strong representation has been extracted.(3)In this model,deep stochastic configuration networks is adopted,which has improved the ability of universal approximation of the classification model.As a result,a classification estimation criterion with strong generalization ability is constructed for the feature vectors of fog image.(4)In order to overcome the problem of poor universality in same feature space,imitating human cognitive estimation model in deliberation and comparison,based on the theory of generalized error and entropy,the paper gives the definition of the indefinitely classification error entropy measure.The real-time evaluation of fog image classification level and visibility to estimate the credibility of the results hasbeen realized.Also,the model is able to dynamically update the number of interleaved group convolutions in VGG16 and the information system granularity,which contributes to further optimize the strong distinguishable ability and the contracted feature maps space,to adjust automatically classification estimation criterion.Finally,the fog images with low reliability are estimated by the transfer learning mechanism.To test the effectiveness of the intelligent classification estimation model,this paper selects 15,000 fog images as a sample library to conduct rank cognition and visibility estimation experiments.The average classification rate of fog perception is 90.23%,and the visibility estimation root-mean-square error is 0.145.The experimental results show that compared with other algorithms,the model has high accuracy of identification,and the estimation ability is also very high.The new model provides the research foundation for the follow-up weather prediction.