Research On Visibility Detection Algorithm Based On Image Processing

As one of the routine projects monitored by meteorological departments,atmospheric visibility is closely concerned by people,especially motor vehicle drivers.Visibility information has attracted much attention in highway transportation,airport aviation,maritime navigation,agricultural production and other fields.Low visibility weather has become one of the main factors leading to traffic accidents.Therefore,it is of great significance to realize the omni-directional detection and real-time prediction of visibility.Traditional visibility detection instruments are expensive,complex to operate and can not be installed on a large scale,which cannot be satisfied.Because scene cameras are widely used,visibility detection based on image processing does not increase hardware cost.Starting from the close relationship between scene image edge information and visibility,this paper chooses visibility detection method of model training as research direction.The Authors use digital image processing technology to extract image features with high correlation with visibility from scene images and form feature vectors.Using machine learning algorithm to establish the relationship model between image feature vectors and visibility truth value,which can be used to calculate the visibility of the image to be measured.Firstly,different schemes for image feature extraction are proposed,and the effects of different schemes on model detection performance are verified and analyzed in subsequent experiments.Secondly,the relevant vector regression machine(RVR)is applied to visibility detection,and grid search algorithm is used to optimize the parameters.Aiming at the visibility detection model based on RVR,this paper mainly studies the influence of different image feature extraction schemes on the performance of model detection,and determines the best scheme through experiments.It is also compared with K-nearest neighbor(KNN),non-parametric optimized RVR and non-parametric optimized support vector regression machine(SVR).The experimental results show that the detection performance of RVR model optimized by grid search algorithm is better than the other three models.Finally,through the analysis of the optimization principles and advantages and disadvantages of genetic algorithm(GA)and grid search algorithm(GSA),aiming at the problems of SVR,such as the large number of parameters to be optimized and the difficulty of optimization,this paper proposes to combine GA and GSA to optimize the parameters of SVR,establishes a visibility detection model based on optimized SVR model,and studies the performance of three different parameter optimization methods through experiments.The experimental results show that the combination of GA and GSA can effectively improve the detection performance of SVR model by optimizing the parameters of SVR model.The performance of RVR model and SVR model is compared and analyzed by experiments.There is no consistency difference in the correlation,accuracy and MAPE between the two models.The detection performance of RVR model in low visibility environment is better than that of high visibility environment,but the training time of RVR model is shorter than that of SVR model.In addition,the single sample test speed of RVR model is faster than that of SVR model,and the model can be better.Applied to visibility detection.