[argyllcms] Re: Possible whitepoint problem

  • From: Gerhard Fuernkranz <nospam456@xxxxxx>
  • To: argyllcms@xxxxxxxxxxxxx
  • Date: Sun, 09 Sep 2007 01:10:35 +0200

If one looks at a K vs. L* plot (see attachment), then it becomes
clearly evident that this device obviously has not been linearized. The
CLUT grid alone can never fit the curvature accurately, in particular a
coarse grid (-ql) cannot. Thus one should not even consider using "-ni"
with this data set; pre-linearization curves are essential here ("-no"
on the other hand does not make so much difference). And using -qm or
-ql instead of -ql helps of course further to fit the data a couple of
0.1 dE more accurately on average. If I additionally replicate the white
patch in the .ti3 file 16 times, then I get

-qh     => WP 96.994025 0.765836 -2.935014
-qh -no => WP 96.976569 0.764789 -2.935626
-qm -no => WP 96.090594 0.746270 -2.901280
-ql -no => WP 95.832088 0.782419 -2.978553

which are all not so much off from 97.510 0.78000 -2.9600.

(normally one should of course not just replicate a single measured
white patch, but rather one should add additional measurements for
white, taken at different locations on the sheet or possibly even from
different sheets (note, a huge target may span several sheets too), in
order that the actual variation can be honored by the profile => targen
-e ...).

The given data set also looks somewhat noisy. I would not expect that
even the -qh profile can predict the color of printed patches much
better than about 2-2.5 dE on average (note, even the self-fit error to
the measurements is about 1.4 dE, and computing a cross-validation also
suggest that decreasing the smoothness is rather not an options for this
data set, but would only fit more of the noise in the measurements,
resulting rather in a worse prediction of the average device behaviour,
than improving things).

I have seen data sets from professional inkjet printers, which did fit
the spline model noticeably better than your data set. The missing
device linearization is indeed one issue (and here is possibly some room
for improvement), but given the already large number of patches, my
feeling is that there might be also a repeatability issue (either of the
printer or the measurements).

Btw, conspicuous is IMO also the rather strong color tint of the black ink:

1 0.0000 0.0000 0.0000 0.0000 97.510 0.78000 -2.9600
32 0.0000 0.0000 0.0000 10.000 77.730 3.0000 -3.8100
33 0.0000 0.0000 0.0000 20.000 60.610 4.7900 -4.9500
34 0.0000 0.0000 0.0000 30.000 47.130 6.4200 -6.2700
35 0.0000 0.0000 0.0000 40.000 35.820 7.5600 -7.6600
36 0.0000 0.0000 0.0000 50.000 27.780 8.0000 -9.0800
37 0.0000 0.0000 0.0000 60.000 23.090 7.1900 -10.030
38 0.0000 0.0000 0.0000 70.000 18.640 6.7800 -10.690
39 0.0000 0.0000 0.0000 80.000 16.300 5.6100 -10.650
40 0.0000 0.0000 0.0000 90.000 15.010 4.4900 -10.490
41 0.0000 0.0000 0.0000 100.00 14.230 3.4100 -10.260


John Weissberg wrote:
> The L value for the wtpt tag varies quite dramatically depending on the
> values used for the q and n tags. This amount of variation seems
> excessive to me.
>
> Here are some values below using various settings for the parameters of
> profile:
>
> profile -v -ql -bl -kr -l333 canon9950
> Lab(wtpt): 92.496330, 0.622382, -3.016244
>
> profile -v -ql -bl -kr -ni -no -l333 canon9950
> Lab(wtpt): 93.403267, 0.365175, -3.695083
>
> profile -v -qm -bl -kr -l333 canon9950
> Lab(wtpt): 89.306440, 0.563566, -2.679620
>
> profile -v -qm -bl -kr -ni -no -l333 canon9950
> Lab(wtpt): 92.164165, 0.582095, -3.329186
>
>
> All calculations performed with Argyll 7.5

PNG image

Other related posts: