Color in Computer Vision: Fundamentals and Applications

Chapter 11: Color Constancy Using Machine Learning

The third type of algorithms estimates the illuminant using a model that is learned on training data. Indeed, gamut-based methods in Chapter 10 can be considered learning based too, but since this approach has been influential in color constancy research it has been discussed separately.

Initial approaches using machine learning techniques are based on neural networks [179]. ¹The input to the neural network consists of a large binarized chromaticity histogram of the input image, the output is two chromaticity values of the estimated illuminant. Although this approach, when trained correctly, can deliver accurate color constancy even when only a few distinct surfaces are present, the training phase requires a large amount of training data. Similar approaches apply support vector regression [180–182] or linear regression techniques such as ridge regression and kernel regression [183–185] to the same type of input data. Alternatively, thin-plate spline interpolation is proposed in Reference 16 to interpolate the color of the light source over a nonuniformly sampled input space (i.e., training images).

11.1 Probabilistic Approaches

Color-by-correlation [187] is generally considered to be a discrete implementation of gamut mapping, but it is actually a more general framework that includes other low-level statistics-based methods such as gray-world and white-patch as well. The canonical gamut is replaced with a correlation matrix. The ...