Colour Constancy: a Simulation by Artificial Neural Nets

Colour of natural objects perceived by humans is almost independent of the illuminant, the phenomenon known as colour constancy. Its mechanisms are not yet completely understood. Calculation of colour-constancy properties is difficult due to a non-linear neuronal activity implied by current models. An alternative is offered by modelling using artificial neural networks. In our study a neural net consists of three receptor types (L, M and S); their signals are transmitted to two opponent (L/M and S/LM) and a non-opponent (Lum) units. From them signals are further conveyed to 'hidden' units, which in turn forward signals to the net output units, selective to the test colour. The net has two channels: one processing information on light reflected from a background and the other from a test. The neural net was trained to identify colour of a set of Munsell samples under various illuminants. It was then tested using original and new illuminants and an extended set of Munsell samples. We conclude that in the neural net identification of the test colour is based on calculation of colour difference between the background and the test. Colour of natural objects perceived by humans is in large degree independent of spectral composition of illuminants. This phenomenon is known as colour constancy (D’Zmura & Lennie, 1986; Lucassen & Walraven, 1993; Troost, 1998; Breive et al, 1999; Malloney, 1999). There are a number of approaches in explaining how the problem of colour constancy may be solved by biological systems. 1. According to the von Kries coefficient model, the gain of three cones is changed in such a way that the cone responses are not altered by the illumination. At present it is not clear how the gain of the photoreceptors is controlled (Malloney, 1999; Wyszecki & Stiles, 2000). 2. Some authors hypothesize that differences in cone responses to light reflected from the background and the objects are independent of the colour of the illuminant (Yarbus, 1975; DeValois et al., 1997). However, experimental or simulation data, which could confirm or reject this hypothesis, lack (Jurkutaitis et al., 2000). 3. Recently Nascimento and Forster (1994) have shown that the ratio of cone responses to light reflected from the surface of different natural objects does not depend on illumination. It is, however, not clear how the ratio is utilized by the visual system to ensure colour constancy. 4. Finally, in accordance with the computational approach, a formal description of processes underlying colour constancy is hypothesized (Hulbert, 1998).