TITLE:Natural scene analysis: implications for cones widths ?

ABSTRACT:
A new analysis of an existing hyperspectral image data-set
of 29 terrestrial natural scenes (Parraga et al, 1998 JOSA
15, 3) may imply a minimum useful chromatic bandwidth for
cones.

As posted at http://www.crs4.it/~gjb/ftpJOSA.html this
data-set is available via ftp. Each scene comprises 31
images. Each image corresponds to a nominal 10 nm waveband,
and the 31 waveband set spans the visible spectrum [400,700]
nm.

The analysis reveals the degree of spatial correlation
between different 10 nm waveband pairs.
As expected, image-pairs of the same scene that correspond
to adjacent 10 nm wavebands are the most similar.
As the separation between the two wavebands increases the
degree of similarity decreases sharply at first, but then
levels off. On average, this levelling-off occurs at around
a separation of 30 nm.

Thus designing a receptor with a half-height spanning less
than 60 nm (two-tail of 30 nm) might not make much sense for
a terrestrial being.  To cover the visible spectrum, such
narrower band receptors would require more cone types but
would gain little in discrimination power - as the images
seen by the new cones could often be rather similar.  This
assumes that the data-set is well representative of natural
scenes, and that receptors optimise to see average rather
than extreme signals.

The analysis - to be presented in greater detail -
first derives reflectance from a greycard reference, then
computes difference images, and finally measures the mean
Fourier energy across a range of chromatic separations for 
various spatial frequency bands.