Research On Single Image Visibility Restoration Algorithm Under Gaussian Model

The rapid development of computer technology and the rise of artificial intelligence have triggered an upsurge of intelligence in various fields,and smart devices have gradually entered into everyday life.As an important branch of computer applications and artificial intelligence,computer vision is gradually applied to image recognition,military surveillance and driverless fields,such as closed-circuit television monitoring systems,airborne surveillance systems,alpha smart driving systems,etc.However,these outdoor working visual systems are subject to weather conditions in real time,and the visual system will be paralyzed due to the scattering,reflection and absorption of light by the medium in the atmosphere.In recent years,China's environmental quality is facing challenges,for example,the hazy weather has caused some interference with the normal operation of the visual system.Meanwhile,haze is a common natural phenomenon,which leads to lower contrast,lower visibility,loss of detail,and severely hindered the image analysis process.Therefore,it is of practical significance and application value to propose some effective sharpening algorithm of haze images.Secondly,from the perspective of research and practice,the cause of haze has a strong regional difference: Southern is dominated by fog or haze,and Northwest is mainly fog or dust haze.And the current algorithm is basically based on the treatment of similar fog/haze environment in the south and does not apply to the fog/dust haze environment in the north,which showing obvious environment dependence.So,designing an image sharpening algorithm under different environments has certain practical significance.For the current image dehazing algorithm,there is a serious environment/image dependence problem.The thesis proposes image clarity algorithms in two kinds of environments.(1)Aiming at the environment of dust haze in the northwest,the theory of optical compensation atomization and the clear algorithm are proposed: Low precipitation in the northwestern region leads to sparse vegetation and loose soil,and dusty weather is very likely to occur.The dust,sand and gravel components in the air are mostly present.The medium absorbs a large amount of blue light and part of the green light,making the image yellow.Through the experimental statistics of the histograms of mist,dense fog and dust haze images,it was found that the histograms of dense fog image corresponding to the three channels of the RGB image are concentrated and distributed in the high grayscale range of [50,220];the histogram of G and B channel of dust image is distributed in the low grayscale range,and the distribution of R channel is similar to that of fog image.Therefore,it is proposed to fit the distribution of the B and G channel of dust image to the R channel,so that it can meet the distribution form of dense fog image.This concept is called optical compensation;besides,the dust image will be classified as fog image by the use of optical compensation,then put forward a pixel-by-pixel concept to estimate the transmission,and obtain the final clear results.(2)Two kinds of defogging methods based on Gaussian model are proposed for the southern haze environment: 1)Iteratively optimized dehazing method based on Gaussian weight attenuation.This method uses Gaussian functions constructed by Kirsch operators to act on haze-free images.Therefore,the concept of Gaussian dark channel is proposed.A large number of experiments have proved that the Gaussian dark channel has a low grayscale distribution range.Next,on the assumption that the transmission is known and can best reflect the image depth information,the prior that the grayscale of the product of the Gaussian dark channel and the optimal transmission in a haze free image tends to zero is proposed to simplify the atmospheric scattering model,resulting in a rough transmission;Finally,a set of high-order filters composed of Kirsch and Laplacian operators are used to iterate rough transmission and obtain the optimal effect.Then we do the backward verification to the above prior,the experimental results prove that the proposed method is established.2)Based on the multi-scale optimized single-image dehazing algorithm,which is inspired by multi-scale Retinex.Three different scales of Gaussian functions are used to obtain the MSR effect.Secondly,the RGB three-channel transmission components and restoration effect are obtained by the atmospheric scattering model.Finally,considering that the current algorithm has a certain image dependence,thus we classify images and establish a corresponding image database.The main items are as follows: 1)According to different characteristics of the fog image,so as to design a targeted clear processing algorithm;2)Using the classification makes further comparison of the current algorithm.3)Based on the classification,the subsequent research directions and ideas can be determined.Based on the above classification,the thesis compares the proposed algorithm with the classical algorithm from both subjective and objective.The experimental results show that the proposed method not only has a subjective effect,but also objectively reflects the algorithm's superiority.