[argyllcms] Re: Hutchcolor targets

  • From: Graeme Gill <graeme@xxxxxxxxxxxxx>
  • To: argyllcms@xxxxxxxxxxxxx
  • Date: Sun, 16 Sep 2007 12:58:39 +1000

dchaffe@xxxxxxxxx wrote:

end seemed okay - they were near 1 and 5.  I tried a regular IT8
target on the same scanner, as well as a reflective target on an Epson
scanner, and got a fairly similar "eye" shape in the blues on a
Granger.  I don't really know enough to judge whether what I see in
the rainbow is a concern or not.

It's hard to say without a lot of work, since there are several complicating
factors. Factors that cross my mind are:

  Examining a chart that setup to look a certain way in one colorspace
  (ie. A rainbow setup in a well behaved additive RGB space), as if it
  was setup for a different colorspace (ie. the RGB space of a real world
  scanner), it's had to know what to expect unless the scanner RGB
  space is similar to the original, and similarly well behaved.
  The whole point of profiling the scanner is usually to deal with the
  fact that it's device colorspace is not so similar, and not so well behaved !

  This is a scanner profile, so the test chart has probably not
  exercised the full device gamut - ie the test patches do not
  take the device output RGB values to their combinatorial limits.
  This means that the resulting profile can only be accurate over
  the RGB gamut that the chart has exercised. In volumes outside
  this, the profiler has necessarily extrapolated (ie. "guessed")
  the CIE input values that would cause such RGB output.

  Your Granger Rainbow is exploring all these out of gamut
  regions, and the artefacts may be partly revealing how
  unreliable the extrapolation is.

Taking up the latter point, you will notice in the resulting
L*a*b* image, that near the blue problems the b* value
is clipped at -128, and in other regions near the input
RGB = 0, 0, 100%, both a* and b* are clipped.

This hints that the explanation might be that the scanners colorspace
is such that if you properly extrapolate from the RGB values exercised by
the test chart out to highly saturated RGB blue values, that you
would get Lab values outside the ICC range (a*b* from -128 to 128),
and naturally these (component wise) clip, creating a strange looking result.
The extrapolation may not be very reasonable compared to real world
behaviour either, exacerbating this problem.
(Argyll's extrapolation is often better behaved than
 some alternate scattered point interpolation algorithms,
 but it knows nothing of real world reasonable behaviour,
 it's just playing with numbers.)
The fact that these anomalies happen at extreme Lab values indicates
that they probably won't trip you up with scanning real world colors.

So there you go, without going into a lot more detailed investigation,
there's my best guess about what's going on.

Graeme Gill.

Other related posts: