[argyllcms] Re: Determining proper error value for -r

Roberto Michelena wrote:

But the key here is to do it only for situations of repeatedly
profiling similar devices.
I agree that it wouldn't make sense for a one-off, but consider this
hypotetical (but quite realistic) scenario:
- I sell proofing systems based on Epson K3 devices and two different
papers (one gloss, one semimatte)
- "relinearization" over a canned (or prebuilt) profile does not cut
it for producing precise matches


I'm surprised to hear that. We (Colorbus) implemented a calibration system
that very successfully kept our proofing systems running within spec.,
without needing to reprofile. It could be that the linearisation
system you are using isn't doing quite the same things I guess.

- so I find myself forced to do a complete profiling of each system I install

now consider this:
1) on my office ("base system") I measure 8 repetitions of a
1000-patch target, which builds a 'reliability' map of the colorspace
2) then I measure an 8000-patch target (standard not repeated patches)
3) I build a "base" profile with this 8000-patch measurements and the
"colorspace tolerance" map
4) I build a link profile (-G) between this "base" profile and ISOCoated
(maybe steps 3 and 4 could be combined)

so now my "base system" in my office has the best possible calibration
to ISOCoated
I have to repeat this for each paper type, of course; so twice.

Then everytime I install a system at a customer, I print and measure a
1000-patch chart through the "base" (canned) simulation, run it
through Refine, and use the produced abstract profile to recreate step
(4) and get a personalized link profile.
Either "refine" or "icclink" would have to incorporate usage of the
reliability map.


It can't possibly apply to icclink, because there is no interpolation
going on there. It could conceivable be applied to "refine",
either in creating the correction mapping or in using a per point
damping factor, but I've had very little feedback about "refine",
and what I did get was rather mixed, so I'm not yet sure if
it does anything useful.

It sounds like you are in an almost unique position
to be able to gather the reliability statistics from a body of printers.
I'm not sure how many other people would be in the same position.

A more interesting intermediate step would be to see if there
is a general pattern to test point reliability in relation
to its position in the colorspace. It wouldn't be at all a surprise
to find that there are systematic effects that are relatively
independent of the particular device. For instance, the instruments
generally use light sensors that have a linear output, and have
some level of noise. When translated into L*a*b*, this may
correspond to there being higher unreliability in dark colors.
Typically printers (no matter what the technology) struggle
to control light tints. This could then translate to a source
of uncertainty that is greater in the light colors.
So if this were the case, it would be possible to apply
a generic test point weighting based on where the point
is in colorspace, that may improve the quality of the resulting profile
without having to go to the trouble of doing multiple
chart readings, or having access to charts read on a sample
set of printers.

Graeme Gill.



Other related posts: