Research On Efficient Dehazing Algorithm For Railway Video

Railway is one of the important transportation facilities in my country.With the increase of transportation demand,the mileage of trains increases year by year,and the speed of trains increases year by year.Therefore,how to ensure the driving safety of the line environment is an important task.By installing a video camera in front of the train,the information of the line environment can be obtained,which provides a strong guarantee for the subsequent safety detection algorithm.Compared with existing datasets,railway video has the characteristics of wide scene,fast driving speed,complex weather conditions and high real-time requirements.However,the quality of the video obtained due to haze weather is not high,and the safety impact on high-speed trains is becoming more and more serious.Therefore,dehazing from railway videos is an important and extremely challenging task.Although image haze removal has achieved good haze removal effects,most of them are for synthetic haze images.In addition,image haze removal is more focused on the modeling of the background bottom layer,ignoring the modeling of the haze layer,the haze layer has certain distribution characteristics in space,and is a continuous image sequence in time,which can be represented by a spatiotemporal process.However,the real railway haze video in practical applications is more complex and has no corresponding clear video,and the existing image haze removal methods cannot solve this problem well.However,the existing video haze removal methods need corresponding clear videos,and it is difficult to solve the problem of real railway video haze removal.For this reason,this paper proposes an efficient haze removal method for railway video,and the research has very important theoretical research significance and practical application value.The main research works of this paper are as follows:(1)Aiming at the problem that real railway haze videos do not have corresponding clear videos and cannot be supervised training,a semi-supervised video haze removal method based on convolutional neural network is proposed.The method restores clear video by modeling haze video frame background bottom layer and dynamic haze layer.Using 3D Markov Random Field Priors to learn the spatiotemporal consistency of the background layers of hazy video frames without corresponding clear videos.Use the haze generator to learn haze videos with corresponding clear videos to generate haze layers,and use the corresponding clear videos as strong constraints.Using two different priors,we design a semi-supervised learning mechanism for synthetic haze videos with corresponding clear videos and real haze videos without corresponding clear videos.Tested on the constructed high-speed railway haze video dataset,compared with existing methods,the semi-supervised haze removal method proposed in this paper can obtain 1.6d B PSNR and 0.01 SSIM gain under mild haze conditions.In the case of heavy haze,the semi-supervised haze removal method proposed in this paper can achieve a PSNR of 1.3d B and an SSIM gain of 0.01.(2)An efficient haze detection module based on machine learning is proposed.The module consists of feature extraction algorithm and machine learning model.The feature extraction algorithm includes grayscale feature extraction、Haar feature extraction and deep learning automatic feature extraction algorithm.Considering the high real-time requirements of the haze detection module,this paper selects the Haar feature with the highest feature extraction efficiency,and the detection algorithm uses the traditional machine learning algorithm.Experiments show that the haze detection module takes 30seconds/hour less time,the accuracy rate is as high as 1.0,and it can make the haze removal module better and more real-time.