Scattering of light due to the presence of aerosol particles along the path of radiation causes atmospheric haze in images. This scattering is significantly less severe in longer wavelength bands than in shorter ones, thus the importance of near-infrared (NIR) information for dehazing color images. This paper first presents an adaptive hyperspectral algorithm that analyzes intensity inconsistencies across spectral bands. It then leverages the algorithm’s results to preserve photorealism of the visible color image during the dehazing. The color images are dehazed through a hyperspectral fusion of color and NIR images, taking into account any inconsistencies that can affect the photorealism. Our dehazing results on real images contain no halo or aliasing artifacts in hazy regions and successfully preserve the color image elsewhere.

[Paper] [Code]

* in authors means equal contributions