Re: Accurate color follow up ...
Re: Accurate color follow up ...
- Subject: Re: Accurate color follow up ...
- From: Klaus Karcher <email@hidden>
- Date: Mon, 06 Jun 2011 13:35:56 +0200
This message apparently did not get through to the list last week (sent:
Fri, 03 Jun 2011 18:20:02 +0200). Therefore I send it once more:
Wheeler, Barry wrote:
THANKS to those of you who have responded to my earlier message. I
am interested in continuing the discussion directly with others who
are interested. In particular, I am interested in exploring camera
profiles and the set of colors used to improve the accurate of
specific cameras.
[...]So I simply import a color set (as
Lab values) into ColorThink and display the colors, a gamut boundry,
and the ColorChecker colors. Thus, by "distant" I am referring to
the visual distance between color points in the ColorThink plots. I
find this presentation is immediately comprehended - and usable! - by
the staff.
The method you describe is indeed useful to estimate whether or not the
gamut of a certain output medium or working space profile is sufficient
to cover the gamut of an original. The gamuts of working space and
output mediums should enclose all possible colors of the original in
order to avoid clipping.
Furthermore the method can be used to estimate whether the gamut of the
profiling target is sufficient to cover the gamut of your originals. As
the extrapolation capabilities of LUT profiles are rather poor in
general, the gamut hull of the profiling target should enclose all
possible colors of the originals in order to avoid extrapolation errors.
So the method you describe is valuable to evaluate two possible sources
of error.
IMHO the pure colorimetric distance of target and original points is not
a very meaningful criterion however: For example you can print a
profiling target that perfectly matches the Lab values of your original.
When you create a profile with this target, you might get very good
results as long as you measure the accuracy of your system with the same
target (or a different target printed with the same pigments) -- but
when you apply this profile to the scan of your original, there might be
still huge and systematic color errors!
A pure colorimetric approach can not reveal the source of this problem
as the source of the problem is "buried" in the /spectral/ properties of
original, target, light source and sensors. From a mathematical point of
view, the tristimulus values (Lab or XYZ values that characterize the
color perception of the human observer) and the RGB values returned by
the camera or scanner are projections of the spectra (in an
high-dimensional spectral space) onto two different 3-dimensional
subspaces. Projections onto lower-dimensional subspaces always imply a
loss of information.
A more descriptive analogy for this kind of projections might be shadows
casted onto projection screens: When you see a circular shadow on a
projection screen, you can not distinguish weather e.g. a ball or a
flat, round disc casted this shadow.
When you use a second projection screen at a different location, the
shadow of the ball will still be a circle, but the shadow of the disc
might degrade to an oval or even to a thin line.
The Lab values we see are just /one/ projection of the "high-dimensional
truth". And the RGB values returned by the Camera or sensor are /a
different/ projection. When you try to build an ICC profile that maps
one of these projections to the other (and exactly this is the role of
the ICC profile), you will get in trouble sooner or later: Neither of
the projections contain enough information to set up an error-free
mapping. Some of the mapping instruction had to be simply contradictive:
The profile has no idea whether the circular shadow steams from a circle
or a disc and therefore it can not map it once to a line and once to a
circle.
<http://digitalproof.info/colorsync-users/paramere.rlc.c3d.73.pdf> shows
measurements of two gray samples. Both samples are almost perfectly
neutral gray and more or less indistinguishable when viewed under D50
(i.e. they are metameric). When you calculate Lab values for both
samples, you'll get almost the same results: Delta E is only 0.76. But
when you capture these samples with a sensor whose red sensitivity is
slightly shifted to the right (and sensors like this are quite common!),
you will inevitably get quite different RGB values.
When you create a profile from a target which contains only the "flat"
kind of grays (that's the case with the ColorCheckerSG for example) but
your original contains pigments which result in curves like the red one
in my plot (curves like this can be found quite often, e.g. in dyed
textiles or photographic prints) the gray patches of the color checker
will be perfectly neutral in the scan, but the scans of many gray
textiles or photographic prints will be way to red.
There are two ways to mitigate this kind of problems:
1) the "projection screen" of the sensor should be as close as possible
to the one for the human standard observer, i.e. the sensor should
satisfy the Luther condition. Unfortunately there are tight constraints
in terms of sensor design and even more in terms of retroactive sensor
optimization and they generally involve hardware modifications.
2) the profiling target should represent the /spectral/ properties of
typical originals as close as possible and the profiling software should
deal with the remaining contradictions as smart as possible.
In order to optimize the color fidelity of Cruse scanners I have
followed both paths. With the old hardware configuration and standard
profiling tools and targets, customers often complained about red shifts
for example (even though e.g. color checker scans were perfectly neutral).
With the modified hardware, my new target and a special profiling
solution (based on Graeme Gill's excellent ArgyllCMS), there are no more
/systematic/ red shifts. As the colorspace in this critical region will
be locally compressed and the profiling software has (and seizes!) the
chance to find the optimal compromise, some pale shades of red and green
are now /a little bit/ too grayish, spectrally flat grays are /a little
bit/ greenish and some grays with high IR reflectance are still /a
little bit/ too red -- but the overall color error is /dramatically/
reduced.
I think in order to quantify the accuracy of reproduction systems
objectively, we badly need standardized, representative /spectrally/
defined test targets. And to optimize the reproduction quality of
cultural heritage materials, we need either multispectral imaging
devices or at least new, smarter ways to incorporate the spectral
properties of these objects in traditional, tristimulus based
reproduction workflows.
Klaus Karcher
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Colorsync-users mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
This email sent to email@hidden