Estimating varying illuminant colours in images

Colour Constancy is the ability to perceive colours independently of varying illumination colour. A human could tell that a white t-shirt was indeed white, even under the presence of blue or red illumination. These illuminant colours would actually make the reflectance colour of the t-shirt bluish or reddish. Humans can, to a good extent, see colours constantly. Getting a computer to achieve the same goal, with a high level of accuracy has proven problematic. Particularly if we wanted to use colour as a main cue in object recognition. If we trained a system on object colours under one illuminant and then tried to recognise the objects under another illuminant, the system would likely fail. Early colour constancy algorithms assumed that an image contains a single uniform illuminant. They would then attempt to estimate the colour of the illuminant to apply a single correction to the entire image. It’s not hard to imagine a scenario where a scene is lit by more than one illuminant. If we take the case of an outdoors scene on a typical summers day, we would see objects brightly lit by sunlight and others that are in shadow. The ambient light in shadows is known to be a different colour to that of direct sunlight (bluish and yellowish respectively). This means that there are at least two illuminant colours to be recovered in this scene. This thesis focuses on the harder case of recovering the illuminant colours when more than one are present in a scene. Early work on this subject made the empirical observation that illuminant colours are actually very predictable compared to surface colours. Real-world illuminants tend not to be greens or purples, but rather blues, yellows and reds. We can think of an illuminant mapping as the function which takes a scene from some unknown illuminant to a known illuminant. We model this mapping as a simple multiplication of the Red, Green and Blue channels of a pixel. It turns out that the set of realistic