[argyllcms] Re: targen options

  Graeme Gill schrieb:

>Are you converting from a device space (e.g. some sort of
>RGB or a CMYK space) ? e.g. How is neutral defined, in terms
>ofd L*a*b*, R=G=B, or K values ?
>
R=G=B in sRGB.

>If so, are you using the same source profile for both the
>former (satisfactory) situation, as well as the unsatisfactory
>one of the RGB inkjet ?
>
Sure, in both cases the source is sRGB, and I also used the same test 
image ("Fujical").

>If not, how sure are you that the source profile is accurate
>in the greys, and that it is not that profiles test chart that
>needs extra grey test patches. (This has certainly been my
>experience in some situations.)
>
sRGB should be pretty accurate in the grays, for a synthetic gray wedge 
with R=B=G.

>Another question/suggestion is on how you supplemented the
>output device test chart. In theory the best way is
>an iterative approach. Profile and link as usual.
>Feed a series of (input space defined) grey values
>into this link, and record the output device values.
>Print and measure them. Add the device and PCS values
>to the output device test chart and re-create the output
>profile. Repeat if necessary.
>
I don't remember exactly, since its a while ago. I think I supplmened 
approx. 125 points near the gray axis, with [a*,b*] values of [0,0], 
[3,0], [-3,0], [0,3], [0,-3], at 25 different lightness levels and 
converted these PCS values to RGB with the old profile, printed and 
measured the patches, and added the results to the .ti3 file. Then I 
created a new profile. But I did not make further iterations (iterations 
are time consuming due to one day drying time for inkjet prints, until 
the color drift stabilizes). I also guess, that my supplemented points 
were distributes too regularly.

>If the above is not the whole problem (ie. lack of adequate sampling
>along the grey axis), then it could be the limitation of the resolution
>of the profile. The basic multi-dimensional representation is of relatively
>coarse resolution (typically 9-33 per axis), and if the device behaviour
>along the neutral axis changes more rapidly than can be represented
>by the multi-d tables, the neutral will not be controlled sufficiently.
>
That's basically what I meant with "corcscrew" characteristics.

>One way of tackling this as to add more points down the neutral axis
>in the test chart (as you have been doing), but also upping the
>resolution of the A2B and B2A tables. This is normally controlled by
>the -q flags, but can be overridden in some of the programs. The cost
>is generally an exponentially greater computation time of course.
>
Sure, but unfortunately even my faster PC (Athlon XP 2600+, not the 
fastest available, but also not so slow) has only a limited computing 
power :-)

>The ultimate answer is that such systems benefit from a calibration
>system. A calibration system would quite finely sample the individual
>colorant channels response, and will quite finely control them. In this
>way the neutral axis can be made almost as perfect as is possible,
>before the profiling system gets to deal with it. (Argyll doesn't
>currently contain tools aimed at this particular task.)
>
Do you mean an (independent) linearization of each channel with high-resolution 
1D luts?

I've no doubts that this is very helpful, if the device channels correspond to
the actual device colorants, since the colorants usually mix rather smoothly. 
This corresponds also to my experience. In my "successful" case, the CMYK 
channels already did have a nearly linear response, when I plotted C,M,Y, or K 
vs. the euclidian CIELAB distance to paper white.

But I'm sceptical, whether this also works well for devices, where the 
controlled channels do NOT directly correspond to the device colorants. My 
inkjet of course prints with CMYK inks, but I can only send RGB values to the 
driver (and I have no a priori knowledge about the driver's internal RGB -> 
CMYK mapping, and the flaws of this mapping). Thus I could only linearize the 
RGB channels. But will linearized RGB channels also imply a smooth RGB->Lab 
mapping near the gray axis? Maybe, if I'm lucky. But maybe not.

Another idea might be to linearize all three RGB channels not separately, but 
together, in order to optimize the grayscale accuracy?

Btw, I've also another CMYK example, which did not work as well as the 
"successful" one. Here the problem was obviously caused by course quantization 
due to a low number of screen levels, particularly in light areas. The profile 
simply outputs 4 values (CMYK), which are quantized independently by the 
halfone threshold arrays, and this also resulted in alternating color shifts 
into C, M, or Y direction along the grayscale (for light grays), if -kh was 
used for profile generation (these color shifts of course disappeared with -kx, 
at the cost of more visible lightness steps, than with -kh, so it's a 
tread-off). A vector quantization in Lab space (to the percetually nearest 
discrete CMYK tuple) would likely give better results in this case, than per 
channel quantization, but of course this would be a too great demand on a 
profile.

Regards,
Gerhard

Other related posts: