[argyllcms] Re: inter-instrument matching tables
- From: Gerhard Fuernkranz <nospam456@xxxxxx>
- To: argyllcms@xxxxxxxxxxxxx
- Date: Thu, 10 Nov 2011 02:17:12 +0100
Am 09.11.2011 20:24, schrieb Roberto Michelena:
What should I use to create the "correction link profile" (Lab to Lab)
when I have the set of two measurements (master and individual) to be
corrected?
As long as it's restricted to one paper/inkset combo, will it be
useful even to uvcut instruments?
I guess that this can be indeed feasible, at least for one particular
paper/ink/printer/... combination.
Good question what's the best model for this kind of correction. One would need to
"play" a bit with the measurement data and evaluate different models in order
to get a feeling which model works how well.
I'd possibly start by investigating a NxM matrix model in spectral space (where
N is the number of spectral bands reported by the first, and M the number of
bands reported by the 2nd instrument), since this approach would even yield
corrected _spectral_ readings (and not just tri-chromatic ones). As the rank of
the reflectance spectra is supposed to be significantly lower than N or M, I
would not derive the correction matrix directly in spectral space, but rather
derive it in a dimensionality-reduced subspace of the spectral space (and
possibly even then some additional regularization may be necessary).
A correction in CIELAB or XYZ space would likely need to be a non-linear
function. Since the inter-instrument difference is (hopefully) small, I'd
possibly start trying a 3x3 matrix model (XYZ space only) and/or 2nd order
multivariate polynomials (for both, XYZ and CIELAB). However, models which
operate in tri-chromatic space (XYZ or CIELAB) are basically only valid for one
particular black generation or separation (I don't have a feeling for the
magnitude of the errors, when the same correction is nevertheless used for
different separations).
Non-parametric, "flexible" models (like splines, RBF, neural networks,...)
could be used as well, given a sufficiently large number of readings, but IMO one must be
careful to limit the number of effective parameters by proper regularization. Equally,
I'd also avoid using polynomials with too high order for this use case.
If you don't want to do any maths or programming yourself, then I think you could possibly
"abuse" the the Argyll "refine" command for your purpose, which creates an
abstract profile (i.e. CIELAB -> CIELAB) from two .ti3 files (what you get here is a non-parametric
smoothing spline model).
Regards,
Gerhard
Other related posts:
