De-hazing of multi-spectral images with evolutionary computing

We developed an algorithm that allows for removing haze from a digital picture by numerically subtracting the contribution of optical scattering by aerosols. The scene is modeled by defining a reflectance function for each pixel, which describes the angular dependence of light scattering at the surface, and by describing the scattering from aerosols with a set of models of varying complexity. An optimization algorithm that mixes downhill methods with evolutionary computing approaches was used to fit the observed image to the model of the scene. The contribution of the aerosol scattering is then removed to obtain a de-hazed image. We present results for multispectral images taken by NASA's Multi-angle Imaging SpectroRadiometer and we discuss the efficiency of the algorithm implemented on a multi-node quadcore cluster computer.