Image Haze Removal Based On Intensity Inversion And Tone Mapping

In fog or haze weather conditions, the visibility of outdoor scene is poor, due tothe atmospheric scattering effect. This not only brings inconvenience to people’s lifeand work, but also makes the captured image degraded severely. With low contrast,distorted color and hidden details, the hazy image is blurred, which greatly reduces itsvalue and jams many fields’ work. Moreover, haze weather in our country is frequentrecently. Study on dehazing has become a hot topic in the field of computer visionand image processing, with profound practical significance. Effective dehazingalgorithm can be applied in areas such as monitoring system and pattern recognition.This paper summarizes the research status of haze removal algorithm, followed adescription of traditional haze removal algorithm using dark channel prior and itsimprovement. The former algorithm adopts physical model and works very well,while the latter uses fast bilateral filtering to accelerate dehazing process and achievesgreat improvement. This paper implements these algorithms and performs fastbilateral filtering on GPU for further acceleration. Then many experiments are carriedout and the strengths and weaknesses of these algorithms are analyzed.The dehazing algorithm based on dark channel prior adopts physical model and itis current state-of-the-art dehazing algorithm. Nevertheless, inspired by Dong et al.’swork[1], this paper proposes a model-free one called iItem (intensity Inverting andTonE Mapping), which is based on the observation that intensity-inverted hazy imagehas similarity with low-light image. It implements tone mapping on intensity-invertedhazy image followed by intensity re-inverting to obtain dehazed result. However,simple intensity inversion plus tone mapping does not bring satisfying result.Therefore, this paper adds maximum-based blocking and bilateral filtering before tonemapping and modifies this low-light oriented tone mapping to achieve successfuldehazing. Experimental results show that iItem is comparable with and even betterthan the algorithm based on dark channel prior in subjective and objective evaluation.