Re: New Topic
Re: New Topic
- Subject: Re: New Topic
- From: "Mark Rice" <email@hidden>
- Date: Thu, 7 Feb 2008 13:56:22 -0500
No, I am not having a laugh. I have used the LVT film recorder, the Durst
Lambda, and the Oce Lightjet. All can deliver precisely calibrated
grayscales for months on end, even though the chemical developing process
may drift. Each value of the gray scale has target densities in the
computer, and the output densities are extremely close to the target
densities.
The Onyx and SAI RIPS, the two most prevalent DO NOT even have any target
densities. The curve shape is recalculated by some form of internal
algorithm (an unknown one that the manufacturers will not release), and is
recalculated each time a re-linearization is done. What the the calculation
based on? Simply ink density limitations, which are selected by eye, by
guess or by golly, by each user. This is the crudest possible process I can
imagine.
To see what a "re-linearization" provides on the average RIP, try this: take
an RGB grayscale and print it in the RIP after linearization has been done,
but before the ICC profile is generated. Everyone I have seen is
ridiculously bad. The reason is the each channel is "linearized"
independently, based on ink limitations chosen via the crude method shown
above. This process is NOT repeatable!
Martin, I did not copyright the process - I simply copyrighted the article I
wrote.
Graeme - you raise some good points.
1. The process does not require inversion, even if the device behavior is
complicated. It simply uses an algorithm that I will compare to herding
sheep - one creates a set of target aim densities, and the iterative process
"herds" the device behavior in that direction. As I mentioned, the iterative
process has to be changed as the targets are more closely approached, or it
will overshoot the target.
2. Again, I have to mention that the starting point for calibration is
critical, and is done in a very haphazard fashion in most RIPS - the
"eyeball" choosing of ink limitations. There is not method of determining
what set of ink limitations produces neutral values on paper.
My frustration is showing because of this problem - I get a calibration that
is nearly neutral, and nearly color appropriate (I was going to say
"accurate", but I knew that would stir up another argument), and when I try
to make it better, it is much more likely to get worse.
Mark
[>] >You're having a laugh aren't you? Or are you working in a different
[>] >"photographic digital imaging field" to the labs that the rest of us
[>] >have to use?
> All photographic photo printers and film recorders use "iterative
> linearization" to obtain precisely controlled grayscales that match
> target
> densities very precisely. I have been begging RIP makers Onyx and
> SAI to
> produce such a RIP, but they don't really see the need for it. My
> problem is
> that if one achieves a linearization that is close, but not dead on,
> one can
> re-linearize, but it is just as likely to make the linearization
> worse as
> better. And, of course, if the linearization changes, the ICC profiles
> become nearly useless.
Pre-press proofing RIPS like the EFI have used iterative calibration
routines to build Lab correction curves for years.
It doesn't follow that iterative calibration is either necessary
or desirable. It doesn't help at all if a one pass calibration
is already at the limit of instrument and machine repeatability.
Calibration and the use of output profiles boils down to
an inversion problem. The device operates in one direction
(forward), and you'd like to invert that behaviour in order
to know what to feed into it to get a certain output.
If the device behaviour is complicated, or the domain
is so large that the forward behaviour can't in practice
be sampled in enough detail or modelled accurately enough
(ie. the 3 or 4 D color behaviour), then an iterative
approach may well be the best one to locate the input
values that produce a target output value. [But note
that you still have a sampling density issue with regard
to how many samples you solve for in the output space!]
For per channel calibration this is probably not the case though.
It's quite practical to sample the channel behaviour in fine detail
and a 1 dimensional (usually monotonic) curve can be very accurately
modelled an inverted. Given a certain patch budget, a better result
and simpler process is probably that of printing a single
detailed test chart that can combine a high sampling density with
some level of noise filtering. Keeping a history and pre-conditioning
the test chart may give a slight increase in accuracy too.
An iterative approach may give less detail between the
target output values, because it wastes it patch budget
testing the same area over and over.
I don't doubt that some (many ?) RIPS don't have a very accurate
calibration system. I suspect the explanation is more likely to
lie in how they set the output targets (are they absolute or relative ?),
how detailed is are the test charts (ie. are they at least 30
patches per channel or more ?), and what sort of algorithms they
are using to model the device behaviour (do they average out
noise at all ?), plus the post calibration device precision used
(if it's only 8 bits per component, the calibration system may
struggle to have enough control).
Graeme Gill.
***********
_______________________________________________
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