Abstract
Natural illuminant and reflectance spectra can generally be well approximated by a linear model with as few as three basis functions. Some models of color appearance further assume that the visual system constructs a linear representation of spectra by estimating the weights of these inferred functions. However, such models do not accommodate nonlinearities in color appearance such as the Abney effect. Previously, we showed that the hue of lights with Gaussian spectra remains constant over much of the spectrum as bandwidth changes, suggesting that the visual system might adopt an assumption like a Gaussian model of spectra so that hue is tied to a fixed inferred property of the stimulus such as the spectral centroid (Mizokami et al, 2006). This model is qualitatively consistent with measures of the Abney effect, and is also consistent with suggestions that natural spectra may in some cases be better described by Gaussian than linear models (MacLeod and Golz, 2003). Here, we examined to what extent this Gaussian inference provides a sufficient approximation of natural color signals. Spectra from available databases, hyperspectral images, and our own measurements were analyzed to test how well the curves could be fit by either a simple Gaussian with 3 parameters (amplitude, peak wavelength and standard deviation) vs. the first three PCA components of standard linear models. The spectra were coded from 400-700 nm in 10nm steps and were fit using the Matlab Optimization toolbox. Results shows that the Gaussian fits were essentially comparable to a linear model with the same degrees of freedom for both reflectance and illumination spectra, suggesting that the Gaussian model could provide a plausible perceptual assumption about stimulus spectra for a trichromatic visual system.