Spectral-based illumination estimation and color correction

We present a statistical technique to characterize the global color distribution in an image. The result can be used for color correction of a single image and for comparison of different images. It is assumed that the object colors are similar to those in a set of colors for which spectral reflectances are available (in our experiments we use spectral measurements of the Munsell and NCS color chips). The logarithm of the spectra can be approximated by finite linear combinations of a small number of basis vectors. We characterize the distributions of the expansion coefficients in an image by their modes (the most probable values). This description does not require the assumption of a special class of probability distributions and it is insensitive to outliers and other perturbations of the distributions. A change of illumination results in a global shift of the expansion coefficients and, thus, also their modes. The recovery of the illuminant is thus reduced to estimating these shift parameters. The calculated light distribution is only an estimate of the true spectral distribution of the illuminant. Direct inverse filtering for normalization may lead to undesirable results, since these processes are often ill-defined. Therefore, we apply regularization techniques in applications (such as automatic color correction) where visual appearance is important. We also demonstrate how to use this characterization of the global color distribution in an image as a tool in color-based search in image databases.