In this thesis, we propose methods for estimation of the colour of the illuminant. First, we investigate the effects of bright pixels on several current colour constancy algorithms. Then we use bright pixels to extend the seminal Gamut Mapping Colour Constancy algorithm. Here we define the White-Patch Gamut as a new extension to this method, comprising the bright pixels of the image. This approach adds new constraints to the standard constraints and improved estimates. Motivated by the effect of bright pixels in illumination estimation, we go on to incorporate consideration of specular reflection per se, which tends to generate bright pixels. To this effect we present a new and effective physics-based colour constancy representation, called the Zeta-Image, which makes use of a novel log-relative-chromaticity planar constraint. This method is fast and requires no training or tunable parameters; moreover, and importantly, it can be useful for removing highlights. We then go on to present a new camera calibration method aimed at finding a straight-line locus, in a special colour feature space, that is traversed by daylights and approximately by specular points. The aim of the calibration is to enable recovering the colour of the illuminant. Finally, we address colour constancy in a novel approach by utilizing unsupervised learning of a model for each training surface in training images. We call this new method Exemplar-Based Colour Constancy. In this method, we find nearest-neighbour models for each test surface and estimate its illumination based on comparing the statistics of nearest-neighbour surfaces and the target surface. We also extend our method to overcome the multiple illuminant problem.
展开▼
