Color by Correlation

Under a large variety of scene illuminants, a human observer sees the same range of colors; a white piece of paper remains resolutely white independent of the color of light under which it is seen. In contrast, color imaging systems (e.g. digital cameras) are less color constant in that they will often infer the color of the scene illuminant incorrectly. Unless the color constancy problem is solved, color appearance models cannot be used to guide image processing, and such processing is necessary for accurate (and acceptable) color reproduction.In this paper we present a new theory of color constancy, Color by Correlation, which solves for the white-point in images by exploiting the correlation that exists between image colors and scene illuminants. For example, because the reddest red camera measurement can only occur under the reddest red light we say that the reddest camera measurement correlates strongly with the reddest light. Importantly all camera measurements correlate to a greater or lesser degree with different colors of light. By examining the correlation between all image colors and all lights we show that it is possible to make a very accurate estimate of the color of the scene illuminant.Color by Correlation not only performs significantly better than other methods but is a simple, elegant solution to a problem that has eluded scientists working on color for over a century.