Research Of Restoration Of Haze Image Based On RETINEX

Bad weather conditions like fog would significantly compromise thequality of the images we take. This would cause bad consequences in thefield of military, traffic and remote-sensing. So it has a great significance tostudy how to enhance the hazy images.This thesis introduces the influence of aerosol particles on lighttransmitting with Fog Degraded Physical Model. The Retinex Theory is aimage enhancement theory based on color constancy. It addresses theproblem of separating the illumination from the original image and to yieldthe reflectance image which is the original appearance of the object.Compared with traditional image enhancement methods, Retinex Theory hasa great superiority for its color constancy, dynamic range compression andsharpening.In this thesis, we firstly introduced the history of Retinex Theory, thenmade a detailed analysis of several methods: McCann Retinex，McCann99Retienx，SSR，MSR，MSRCR and Variational Retinex. With the results ofexperiments, we evaluated advantages and disadvantages of these traditionalmethods regarding improving image quality, color consistency andcomputation speed. On the basis of illuminance image segmentation and mathematical analysis, we made some modifications on Retinex algorithms:(1) By using logarithmic image processing models image segmentation,the original image is segmented into four regions with different illuminancefeatures. Then a SSR scale parameter is decided according to the illumiancedifference. The enhanced sub-images are fused based on the proportionalityfactors to achieve image enhancement. Illuminance Segmentation Retinexcan reflect the merit of each scale in color maintain or detail enhancement.(2) Based on mathematical analysis, Path-Wise Retinex Algorithmshave been proved to be redundant. Path-Wise Retinex Algorithms have todeal with the following problems: strong dependency on paths geometry,high computation cost and sampling noise. We introduced a newimplementation---Random Set Retinex. In this theory we used2-D pixel sets,such as areas, instead of using1-D paths to analyze locality of colorperception. Random Set Retinex parameter’s tuning can be accomplished byan unsupervised method based on quantitative analysis. In addition, RandomSet Retinex can get a faster performance than other Retinex algorithms.