[argyllcms] Re: Determining proper error value for -r
- From: Gerhard Fuernkranz <nospam456@xxxxxx>
- To: argyllcms@xxxxxxxxxxxxx
- Date: Fri, 26 May 2006 23:09:14 +0200
Ben Goren wrote:
A low -r value (but, of course, not so low as to have real trouble
with noise) may produce more accurate colors at the expense of
smoothness in photographic images. A high -r value will result in
colors that're less accurate, but perhaps not so inaccurate as to be
noticeable outside of a side-to-side comparison. I'm pretty sure
from some preliminary testing, though, that too high an -r value
will start to cost you fine detail, so it may remain something
that needs to be tweaked individually.
Ben, I'm objecting the statement "a low -r value will produce more
accurate colors". That's basically not the case. A low smoothness factor
will indeed result in a model which fits the _training set_ with a low
error. But is that an accurate profile? Such a profile fits the noise in
the measurements. It will only be accurate for the single (already
elapsed) print job, which did print the target, but it won't be just as
accurate for future print jobs, where the printer is expected to print
small random variations of the colors due to its reproduction error, and
where the print job may contain completely different colors than the
previously printed training set. Since the profile cannot eliminate the
reproduction error of the device, the best thing it can do, is to
attempt to characterize the _average_ device behaviour as accurate as
possible. And regarding this accuracy goal, it is _not_ the lowest
smoothing factor which gives the best accuracy, but there exists indeed
an optimal smoothness factor (not too low, and not too high), for which
the error of the model with regard to the average device behaviour
becomes minimal. Only if you increase the smoothness beyond this optimal
point, you're starting to trade-off accuracy against smoothness. If you
go with the smoothness below this point, then in fact you're sacrificing
both, smoothness _and_ accuracy :-(
Regards,
Gerhard
Other related posts: