Research On Daytime And Nighttime Image Dehazing Algorithms

With the rapid development of technology and industrialization,hazy weather has become a very common natural phenomenon.Atmospheric suspended particles such as water droplets and particulate matter will scatter the atmosphere light,resulting in a series of image degradation such as low contrast,unclear details,high local brightness,reduced saturation and other degradation phenomena under hazy condition.The image degradation has a great impact on the subsequent technical work in the field of computer vision such as target detection and tracking,image recognition,etc.,and brings great inconvenience to the work of medical,military,transportation and other work of industries.Therefore,it is of great importance to study how to recover hazy images,improve the clarity of images under hazy conditions,and reduce the impact of hazy weather on the images acquired by imaging systems.Most of the current dehazing methods can deal with daytime outdoor hazy images,and are based on the assumption of uniform distribution of atmospheric light,which is not applicable to nighttime hazy images.Due to the active light sources,such as street lights and headlights of motor vehicles,the nighttime hazy images often contain a large number of dark areas,which makes many details of interest more difficult to distinguish.In order to solve the above problems,this thesis analyzes the principle and model of the image degradation in detail,investigates the basic theories and the key techniques of image dehazing in depth.Some meaningful practical work has been done in terms of perfecting and improving the existing image dehazing methods and new ideas have been introduced.The specific research contents are as follows:1.For the problem of unrealistic sky regions recovery in daytime dehazing,the sky region segmentation method is proposed in this thesis.The basic idea is to segment the sky region by using two constraints of gradient and luminance,and to obtain the probability distribution of the pixel belonging to the sky regions.This method is able to identify and segment the sky region,which is helpful to improve the recovery of the sky region and lays the foundation for the research of improving the precision of atmosphere light and transmission estimation.2.To address the problem of low accuracy of haze density estimation methods,a haze density model related to haze-relevant features is proposed in this thesis.Since the haze density is related to the local brightness,saturation and gradient of the pixels,a linear model can be used to describe the relationships with characteristics.After obtaining the rough estimation map of the haze density,the map of haze density is smoothed by using guidefilter,and the results improve the precision of haze density in detail.3.To solve the problem of artificial light source halo with large area in nighttime hazy images,the image layer decomposition method is improved to remove glow and correct color shift.The glow image is separated by the method of image de-Gaussian blur combined with white color-balanced according to the gradient histogram feature,so that the separated image light sources shape remains intact and the color looks more natural.4.Focusing on the problem of inaccurate estimation for atmosphere light and transmission in bright and dark areas of nighttime hazy images,the method to segment the light source area based on the difference of color channels is proposed,and this channel difference is regarded as the probability of the pixel which belongs to the light source areas.In this way,the atmosphere light and transmission of light and non-light source areas can be estimated separately,and the probability value can be superimposed to improve the estimation accuracy of the atmosphere light and transmission.5.For the difference of image transmission between the light source and non-light source regions,the depth of the image is firstly estimated by counting the information of illumination,saturation and gradient.It is found that they are positively correlated with the image depth,so a nonlinear depth estimation model can be developed to obtain the transmission value.The transmission in non-light source regions is then estimated by using the dark channel prior.Finally,the transmission of the whole image can be obtained according to the probability of the pixel which belongs to the light source regions.6.To verify that the proposed algorithm has significant dehazing effect,this thesis utilizes image quality assessment parameters such as structural similarity,peak signalto-noise ratio,contrast gain,and image blind assessment to evaluate the experimental results from both subjective and objective perspectives.In addition,the running time of the algorithm is recorded and compared with other real-time algorithms in this thesis.According to the experiments,it can be seen that thedehazing algorithm in this thesis runs fast and removes the haze significantly.