Research On Image Haze Removal Based On Airlight Veil And Convolutional Neural Network

Bad weather such as rain and haze could cause image degradation that token outdoors.The application system based on computer vision technology is widely used in various fields,recently.However,these application systems cannot work effectively with the blurred image that affected by the haze,because almost all of these systems work based on clear image.The blurred image has lower value than clear image so that it significantly to do research on dehazing.It is difficulty to dehazing because it cannot know the clear image if it would not effect by bad weather,so that dehazing is uncertain.Generally,dehazing method add some limitations and build up a model to simulate how the bad weather affect image.However,these limitations is not so exactly that dehazing method can be improve.This paper is to do research on dehazing for the blurred image.First of all,doing research on atmospheric scattering model and propose a dehazing method that self-adaptation dehazing(SAD)method based on atmosphere veil.SAD method can limits the "Halo" effect that always appear in the bright area of the image such as sky area.Moreover,doing research on convolution neural network and propose a dehazing method based on ResNet(DoR)to overcome the limitation of dehazing method which based on atmospheric scattering model.DoR method is to find the relationship between clear image and hazy image,and then creating clear image by the hazy image with the relationship.With the current theories,this paper studied as follow:(1)To limit the "Halo" effect and protect white scene in the image while dehazing,proposing a dehazing method that self-adaptation dehazing(SAD)method based on atmosphere veil.To verify the objective validity and real-time of the SAD algorithm,doing some contrastive experiment with other state-of-the-art dehazing method such as He's method and Tarel' method.Firstly,SAD method build up a low pass filter with Laplasse operator to retain edge;and then build up a coefficient to correct the atmosphere veil with the saturation.Finally,SAD method use air-light map to restore hazy image while current dehazing method regarding air-light as a constant value.Experiment result shows that SAD is faster 30% than He's method,170% than Tarel's method and 9% than Zhu's method;also SAD do better than others in MSE and comprehensive evaluation index.(2)Though SAD do good at dehazing but is also has limitation that it cannot process hazy image for different scene.To overcome this problem,proposing a dehazing method that based on ResNet(DoR).DoR build up relationship between clear image and hazy image with ResNet.Using image from different scene to train this network,can lead to the network learn how the relationship is between clear image and hazy image.The network designed end-to-end so that it avoid human interference so the relationship is accurately.The experiment result shows that DoR method do better than current state-of-the-art dehazing method and SAD method.