Color constant descriptors

Land (1983, 1986b) Land et al. (1983) also showed that the visual cortex is involved in color perception. Originally, it was not known whether the computation of color constant descriptors is carried out inside the retina or the visual cortex. Marr (1974) suggested in 1974 that lightness computations are carried out inside the retina. He gave a detailed... 

For the latter case (computation of colors under a canonical illuminant Lc), color constant descriptors are given by... 

The human retina contains only three types of receptors, which respond mainly to light in the red, green, and blue parts of the spectrum. Suppose that we equip a digital camera with filters that model the response characteristics of the receptors found in the retina. If the processing done by the retina and by the brain were completely understood, we could compute the same color constant descriptors as are computed by the human brain. This method could then be used for object recognition based on color. 

From this experiment, we see that the surround does have an impact on color perception. If only a single patch is illuminated brightly, then this patch is assumed to be white as no other color information is available. However, if the surround is also illuminated then the viewed scene becomes sufficiently complex such that color constant descriptors can now be computed. 

Figure 11.10 Computation of color constant descriptors using normalized color shifts. The first layer is the input layer. The second layer computes local space average color. The output is computed in the third layer. The measured red, green, and blue components are stored in cells c, with i e r, g, b. The average of these three components is stored in cell c. Space average color is stored in cells a,, the average of these three components in stored in a. Output cells are denoted by n, (from Ebner 2003c).

Figure 14.2 Likely description of the color constancy algorithm used by the human visual system to compute color constant descriptors. The parameter p is simply a small percentage as described in Section 10.2.

Gershon R, Jepson AD and Tsotsos JK 1987 From [R,G,B] to surface reflectance Computing color constant descriptors in images In Proceedings of the Tenth International Joint Conference on Artificial Intelligence, Milan, Italy (ed. McDermott JP), Volume 2, Morgan Kaufmann, pp. 755-758. 

Some authors also take the approach to compute images that look as if they were taken under some canonical illuminant (e.g. Forsyth (1992)). In general, it is important that the output computed by a color constancy algorithm is not affected by changes of the illuminant. In other words, the computed descriptor should be constant irrespective of the illuminant. Otherwise, we would not have a color constancy algorithm. But at the same time, the output has to be nontrivial. If we would not make this assumption, a trivial color constancy algorithm could be found, which returns the same constant value for all input images. This was also pointed out by Forsyth (1992). 

---