Re: arbitrary profiling targets
Re: arbitrary profiling targets
- Subject: Re: arbitrary profiling targets
- From: Graeme Gill <email@hidden>
- Date: Sun, 21 May 2006 22:59:20 +1000
Roger Breton wrote:
I would be very curious to hear from someone who actually understands the
math.
Maybe Grame Gill could step in the discussion?
I can't help a whole lot more than other people on this list,
as to what number of patches is too many to make a visible difference
in practice.
My personal experience with some large format printers and
a particular RIP system that incorporated a good calibration system,
running in CMYK mode, was that the difference between 1000 patches
an 3000 was visible. Beyond that, improvements were more subtle.
Now, with a different device, different test chart test generation
algorithm, no calibration system etc., the results might be different,
and maybe more test points may give a better result.
The underlying math helps in understanding things, but it depends
heavily on the algorithm used in creating the regular grid table
that a LUT based ICC profile contains.
The statistically correct modelling approach takes into account
that the measured values are not perfectly accurate, but incorporate
sources of error. If the statistical nature of the errors is
known (the level of inaccuracy, and the distribution of
errors), then the model can allow for this, and try and
produce the "most likely" underlying profile. Expectations
of device behaviour (if known accurately) can also help in
this process.
In general, if the above modelling process is being used,
there are two factors that affect how many patches are
needed to achieve a certain accuracy. One is the
underlying behaviour of the device, i.e.. how the
color changes with device values. If there are
areas in the colorspace that change quickly, or
not very smoothly, then dense sampling is needed
in those areas to map the behaviour. Insufficient
density of sampling is analogous to aliasing errors.
The other effect is random errors. The greater the level
of errors ("noise" level), the greater the number of samples
needed to average the noise out, and arrive at the underlying
behaviour.
Now for the latter consideration, printing and reading the same
chart several times will probably have a similar effect, but
if you're going to go to that trouble, it's actually
cleverer to use the same number of total printed patches
to sample more densely, at the same time.
If the modelling process is not properly taking account
the error statistics in the patch values, or doesn't
have sufficient information about the error statistics
(it is not "filtering" out the noise sufficiently), then
in fact more patches may lead to a profile with worse underlying
accuracy in some situations.
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