Re: some thoughts on CIELAB, part 1
Re: some thoughts on CIELAB, part 1
- Subject: Re: some thoughts on CIELAB, part 1
- From: Graeme Gill <email@hidden>
- Date: Mon, 10 Feb 2003 18:52:32 +1100
Robin Myers wrote:
>
The hooking makes a difference if the programmer writing the profile
>
generation code uses simple extrapolation code for taking the target
>
measurements and extending them for creating a larger sampling than the
>
target provides. By not understanding the "hooking" and the curvature of
>
constant hue lines the resulting profile creates purple skies instead of
>
blue ones. I have experienced this problem with many clients and several
>
different profile generation packages.
Extrapolation certainly works better if you are using a model based profile
that encompasses the essence of a devices behaviour. I agree that table
based profiles don't generally do extrapolation very well, and quickly
loose accuracy outside the gamut of the measurement points they
are based on. Presumably the only devices where this should be an
issue for are input devices. I would think that this has more to do with
the fundamental assumptions of model vs. table profiles, rather than
the particular choice of colorspaces. On the other hand, a model
based profile can give terrible results if the device in question
has a fundamentally different underlying mechanism to what the model assumes,
and can give poor accuracy if the device has real world secondary
behaviour effects (ie. non-additive behaviour according to the
models view) that are note allowed for in the model.
Most of the "blue turns to purple" situations I've come across can be
attributed to doing gamut mapping in L*a*b* space, where a way out of
gamut blue is being mapped into a small destination gamut.
>
I'm not sure what you mean by "useful" here. I have used XYZ mixing
>
algorithms with all manner of color production devices quite
>
successfully. By successful I mean that within the gamut of the device I
>
have achieved colorimetric reproduction to within the error of
>
production of the device. The technologies I have successfully profiled
>
this way include CRT, LCD, thermal transer, dye diffusion thermal
>
transfer, printing press, ink jet, hot wax ink jet, electrostatic liquid
>
toning and electrostatic solid toning. The algorithms used are
>
documented in a US patent 5,909,291.
I understand that this sort of approach can work well, but isn't it true that you
had to tune each model for the peculiarities of each type of device technology ?
(No I'm not going to go and read the patents, since most patents seem to go
out of their way to claim everything while revealing nothing, and
idea and algorithm patents simply retard the the progress of science.)
How well for instance, would you imagine that your printing press model
would cope with a CMYK "device" that is actually a fake CMYK space
that gets separated into CMYK + light CM + Orange + Green by
a subsequent process ? A table based profiling approach, while less
elegant and than a model based approach, is a reasonable "one size fits all"
compromise.
>
The output device ignorance in the current ICC model means that often
>
gamut mapping is performed when it is totally unnecessary (i.e. when the
>
source color is within the gamut of the destination device) or that
>
intelligent gamut mapping (using the source and destination gamuts to
>
adjust the amount of mapping) is extremely difficult, if not impossible.
Agreed. I've been pointing this out for some time. As for intelligent
gamut mapping using ICC, it is perfectly possible, but not fast. My
CMS for instance will customize the gamut mapping during linking for
the source and destination color space gamuts, or even the source
gamut of the image to be transformed. The "secret" is to ignore the
B2A tables and use the absolute A2B (native device response) information
only.
>
There are at least two sets of perceptual "standards" that I know of,
>
Munsell and OSA. Unfortunately, I do not have a set of XYZ values for
>
the OSA standard and so I have used Munsell for basing my analysis of
>
L*a*b* space.
I've seen reference to six data sets being used for CIECAM97 evaluation,
CII-Zhu, OSA, Guan, BFDB-Textile, BFDB-Paint, and Munsell. See the
document "A uniform colour space based upon CIECAM97s" by
C. Li and M. R. Luo at <
http://www.colour.org/tc8-01/Li-final.doc>
if you would like more details.
Graeme Gill.
_______________________________________________
colorsync-users mailing list | email@hidden
Help/Unsubscribe/Archives:
http://www.lists.apple.com/mailman/listinfo/colorsync-users
Do not post admin requests to the list. They will be ignored.