Dimensionality of color space in natural images

The color histogram (or color cloud) of a digital image displays the colors present in an image regardless of their spatial location and can be visualized in (R,G,B) coordinates. Therefore, it contains essential information about the structure of colors in natural scenes. The analysis and visual exploration of this structure is difficult. The color cloud being thick, its more dense points are hidden in the clutter. Thus, it is impossible to properly visualize the cloud density. This paper proposes a visualization method that also enables one to validate a general model for color clouds. It argues first by physical arguments that the color cloud must be essentially a two-dimensional (2D) manifold. A color cloud-filtering algorithm is proposed to reveal this 2D structure. A quantitative analysis shows that the reconstructed 2D manifold is strikingly close to the color cloud and only marginally depends on the filtering parameter. Thanks to this algorithm, it is finally possible to visualize the color cloud density as a gray-level function defined on the 2D manifold.