Page 177 - Computational Colour Science Using MATLAB

164 MULTISPECTRAL IMAGING
colour constancy. The problem of whether the visual system might be able to
recover the spectral properties of objects in a scene from the cone excitations has
been studied extensively and analyses of this problem are relevant for
multispectral imaging. We describe some computational procedures for spectral
recovery using multispectral imaging and finally describe some applications of
these procedures for reflectance recovery and camera characterization.

10.2 Computational colour constancy and linear models

Colour constancy, the phenomenon by which surfaces tend to retain their
approximate daylight colour appearance when viewed under a wide range of
different light sources, was described in Chapter 6. It is still a mystery how the
visual system is able to discount the effect of the illumination when the colour
signal that reaches the eye depends just as much on the spectral power
distribution of the illuminant as it does on the spectral reflectance of the surfaces
in the scene (Hurlbert, 1991). One possible mechanism that could account for
colour constancy is adaptation of the light receptors or cones. Such a possibility
was first put forward by Von Kries and is consistent with the chromatic-
adaptation transforms that were described in Chapter 6. However, adaptation is
a relatively slow process and yet colour constancy seems to occur almost
instantaneously as we move from one light source to another in our everyday
lives. An alternative approach to adaptation was postulated by Land and
McCann (1971) who suggested that the visual system may use some
computational process to recover signals that are independent of the illumination
in a scene. In their computational analysis, known as the Retinex theory, Land
and McCann called these signals lightnesses, biological correlates of reflectance
that were computed by each of the three channels in the visual system. Later, the
term integrated reflectance was introduced (McCann et al., 1976) to describe the
illuminant-invariant signals. A number of researchers have since investigated to
what extent the visual system might actually be able to recover spectral
reflectances for points in a scene from the corresponding triplets of cone
excitations.
There are serious limitations on what we can achieve when we set out to
estimate surface reflectance from cone excitations. For example, theoretically
there is an infinite number of combinations of the surface-reflectance functions P
and illuminant power distributions E that could produce a given colour signal S.
In addition, the visual system does not measure S directly, but rather it encodes
the absorption rates of the three different cone types. It seems that if P and E
were not constrained in some way, then the cone excitations would provide little
useful information; fortunately there are some strong constraints on both P
and E.