Foggy And Hazy Visibility Estimation Method Based On PM-DiracNet

The continuous development of economy and science and technology usually brings increasing environmental pollution.With the continuous improvement of China’s industrialization level,the degree of haze in cities is gradually deepening,and haze will seriously disrupt people’s visibility.Low visibility not only affects people’s daily lives,but also more likely to cause traffic accidents,thereby endangering people’s health.Therefore,how to improve the accuracy of visibility detection has become a difficult problem that must be solved now.In view of the many disadvantages of the traditional visibility detection algorithms used by previous researchers,such as low detection accuracy,inability to detect in real time,and low accuracy for small sample data.In order to overcome the limitations of the traditional visibility algorithm,this paper proposes a haze visibility detection algorithm based on the haze scenes in the city and combined with the relevant knowledge of deep learning.The main research contents are as follows:First,in this paper,haze data collection points are set up in large cities,and the haze visibility data set is established through data preprocessing.Based on this data set,ResNet and DiracNet are used to construct the relationship between the haze image and the visibility value,and the comparison baseline(Benchmark)of the algorithm proposed in this article is established,that is,the ResNetbased haze visibility algorithm(Benchmark-I)and the DiracNet-based Haze visibility detection algorithm(Benchmark-II).Secondly,this paper analyzes the advantages and disadvantages of ResNet and DiracNet convolutional neural networks,proposes an improved DiracNet convolutional neural network,and applies it to haze visibility detection.A visibility detection algorithm based on improved DiracNet is constructed.Finally,the above data set is used for algorithm verification and compared with Benchmark-I and Benchmark-II.The results of the test data show that the mean absolute percentage error(MAPE)value obtained by the improved visibility detection algorithm of DiracNet is 2.24%,and the Benchmark-I and Benchmark-II have MAPE values of 5.72% and 4.73%,respectively.The algorithm verification results prove the effectiveness of the algorithm and Superiority.Finally,this article focuses on solving the problem of overfitting caused by training of small sample data,which leads to the problem of low accuracy of visibility detection.Through research and analysis,we found that the concentration of fine particles in the atmosphere(PM1.0,PM2.5,PM10)plays an important determinant of the degree of smog pollution.Therefore,this paper combines PM values with transfer learning to build a PM-DiracNet-based Visibility detection algorithm and experimental verification based on smog data set.The verification results show that the visibility detection algorithm based on PM-DiracNet can greatly improve the accuracy of visibility detection under small sample data,and can reduce the MAPE value from 20% to less than 12%,which proves the effectiveness and superiority of the algorithm.