Image Classification Of Highway Cluster Fog Based On Machine Learning

Since the reform and opening,our country is developing rapidly in all walks of life.Moreover,it has been a steady rise in the national economy.The development of most industries is inseparable from our country’s increasingly convenient transportation.In other words,the development of transportation industry is the foundation of the economic development of our country.Meanwhile,highway as an integral part of the traffic industry,its importance is self-evident.Highway safety work also has always been an important task of traffic department in China.There is a potential safety hazard to solve tricky in highway safety work,which is the phenomenon of highway fog.The fog on the highway has strong suddenness,small scope and unpredictable.This usually leads to huge casualties.Because the hidden danger of highway fog is great and difficult to prevent,universities and companies in Province jointly set up a major special project--Highway mass of fog detection based on Beidou navigation key technology research and application demonstration,which funded by the department of science and technology of Anhui Province.The purpose is to solve the problem of detection and early warning of highway group fog.This paper belongs to its sub-project,which the main research content is the classification of highway fog images.Machine learning has made a considerable number of achievements in the field of image classification.In the paper,machine learning classification algorithm is used to classify highway fog images.If the number of classification categories is two,the highway condition contains cluster fog.At present,there is no open source data set for highway fog images.The data set in the paper is constructed by taking pictures of the videos taken by the on-board driving recorder of car passing through highway fog.The first and foremost is to use traditional machine learning algorithms,including K nearest neighbors,support vector machines and BP neural network algorithms.Three groups of features are selected for the feature extraction of images.The first group is the basic gray features of the image,including the average value,variance,standard deviation,median value and so on.Group two for image Haar feature,which can well show the image gray level change.The third group is the gray value feature of image pixel.Namely,the feature is not extracted artificially.After pretreatment of image pixel gray value are all as characteristics.The way of parameter selection of three different classification algorithms is also different.KNN with exhaustion method,PSO algorithm is used to select parameters of SVM and BP neural network initialization of a set of parameters,which based on loss function and the number of iterations each experiment images for parameter adjustment.Experiment with three traditional machine learning algorithms,selecting the optimal parameters of cases,respectively on the test set has achieved 87%,89%,90%accuracy.This paper uses the deep learning method for the highway fog image classification experiment after using the traditional machine learning algorithm for experiments.Deep learning is a branch of machine learning,which has developed rapidly in recent years.It has been applied in many projects and achieved excellent results.Deep learning in the direction of image classification is the main application of convolution neural network.This paper first to the expansion of image data sets,then design the network model.Mini-batch gradient optimization algorithm is used in order to speed up the updating of network model parameters.Adam optimization algorithm and early stopping are used in order to speed up the training speed and make the loss function converge accurately.The connection layer as well as the convolution model layer,respectively,to join the dropout and BN to prevent model fitting.The model uses precision,recall and F1 value to evaluate.In order to make the model more robust,K-value cross validation method is used to divide the data set.Finally,the precision,recall and F1 value of the data set are 94.2%,93,9% and 94% respectively.Figure [34] table [12] reference [49]...