Haze removal with channel-wise scattering coefficient awareness based on grey pixels.

Abstract

Before being captured by observers, the information carried by light may be attenuated by the transmission medium. According to the atmospheric scattering model, this attenuation is wavelength-dependent and increases with distance. However, most existing haze removal methods ignore this wavelength dependency and therefore cannot handle well the color distortions caused by it. To solve this problem, we propose a scattering coefficient awareness method based on the image formation model. The proposed method first makes an initial transmission estimation by the dark channel prior and then calculates the scattering coefficient ratios based on the initial transmission map and the grey pixels in the image. After that, fine transmission maps in RGB channels are calculated from these ratios and compensated for in sky areas. A global correction is also applied to eliminate the color bias induced by the light source before the final output. Qualitatively and quantitatively compared on synthetic and real images against state-of-the-art methods, the proposed method provides better results for the scenes with either white fog or colorized haze.