Automatic Multi-light White Balance Using Illumination Gradients and Color Space Projection

White balance algorithms try to remove color casts in images caused by non-white scene illuminants, transforming the images to appear as if they were taken under a canonical light source. We propose a new white balance algorithm for scenes with multiple lights, which requires that the colors of all scene illuminants and their relative contributions to each image pixel are determined. Prior work on multi-illuminant white balance either required user input or made restrictive assumptions. We identify light colors as areas of maximum gradients in the indirect lighting component. The colors of each maximal point are clustered in RGB space in order to estimate the distinct global light colors. Once the light colors are determined, we project each image pixel in RGB space to determine the relative contribution of each distinct light color to each image pixel. Our white balance method for images with multiple light sources is fully automated and amenable to hardware implementation.