Scanner profiling issues

Help
peto
2009-01-17
2013-05-01
  • peto
    peto
    2009-01-17

    I'm trying to do (actually, help someone do) a straightforward scanner profile, and the results aren't coming out too good:

    lcms profile ("huge" resolution, linear bradford, default dev 0.005, smoothness 1.0): http://img164.imageshack.us/img164/4498/profilelcmsub0.jpg
    no profile (source for the above): http://img142.imageshack.us/img142/6182/profilenoneij6.jpg
    calibration target scan (if anyone wants to try it themselves): http://img187.imageshack.us/img187/6100/sgsmallxa5.png
    manual color adjustment: http://img50.imageshack.us/img50/4376/manualuq8.jpg

    I don't have the original media to compare against, but he describes the original as being extremely high-contrast, and even without that description the LCMS version looks pretty washed out.  (Auto-contrast in Photoshop helps a bit, but the reason for the calibration is to avoid things like that--and the colors still don't look as good as the manual one.)  Other images have similar results.

    Using a CanoScan 5600F.  All scans are 24-bit, from VueScan with no color balance adjustment.  (For some reason, it seems like this scanner's drivers only do 48-bit with TWAIN drivers; the WIA drivers seem to always scan in 24-bit, and VueScan doesn't support TWAIN.  At least, that's the current diagnosis.)

    There's some messiness in the target scan in the A row (described as lamination), but it doesn't seem to be affecting the profiling; tweaking the alignment to avoid it doesn't make any difference.

    White point near  D50
    Media white (XYZ): 107.47, 111.09, 92.50
    Primaries: R:0.71, 0.26  G:0.31, 0.69  B:0.083, 1.7e-05
    Estimated gamma: R:2.34, G:2.16  B:2.
    Average dE : 1.33
    Standard deviation: 1.51
    Peak: 9.68
    Min: 0.07
    Average Target error: (99% confidence) 0.90
    Maximum error induced by target itself: 1.68

    I know there are a lot of things that can be done wrong while profiling, but we've fiddled with this for a while now--so maybe it's a common problem that someone here will recognize.

    I'd like to try CVS, but building it in Windows is a bit of a nightmare (after about half an hour convincing it to find the simple image libraries, when it suddenly told me it wanted the WinDDK too I threw up my arms and quit).  It'd be nice if someone would drop a current build somewhere (or if the build process in Windows could be cleaned up).  I'll put more time into trying to build it if anyone thinks it'll actually help here, though.

    (I'm glad I copied this message elsewhere on a hunch before submitting, as it happily told me I was no longer logged in and discarded the whole thing, and it was lost from history.  Stupid SF.)

     
    • IMO everything looks pretty reasonable.

      You have just made the following experiences:

      (1) An exact reproduction of the same color stimulus in different viewing environments (e.g. reflective print versus monitor, or versus original outdoors scene, etc.) does NOT IMPLY that the subjective color appearance would be the same as well. For instance, a colorimetrically correct reproduction of a captured real world scene or a reproduction of a reflective print on a monitor in a dim environment will always appear more or less washed out, even if the color stimulus (relative to monitor white or paper white) is reproduced exactly.

      (2) The contrast of the printed medium you have scanned is in fact pretty low, compared to the contrast that can be reproduced on a monitor. The "black" on the print simply is not darker than it gets reproduced by the color managed scan on a [well calibrated] monitor.

      Nevertheless, when you view the print directly, your vision adapts to this low contrast and is happy [due to your congnitive experience that reflective prints generally do not have a higher contrasts], however if an image with the same low contrast ratio is displayed on a monitor, your vision is no longer happy with the reproduction and will judge it as washed out, since it's your experience that the monitor can reproduce much larger contrasts. This is not an issue of a correct or incorrect reprodcution of the color stimulus, but this adaptation happens in your brain.

      And yes, if you want to make the subjective component of your vision happy as well, then you need to stretch the contrast additionally; an exact colorimetric reproduction is not sufficient in order that you get the appearance you expect [actually the color stimuli need to be reproduced _wrongly_, in order to make the subjective component of the human vision happy].

      The scanner profile, if applied with relative colorimetric intent will only take care of a colorimetrically correct reproduction of the stimululs.

      The perceptual intent of profiles created with the CVS version of lprof will additionally take care of issue (1) mentioned above , granted that you supply the correct CIECAM02 viewing conditions for the source and the destination viewing environments. It still does not address issue (2).

      Regards,
      Gerhard

       
    • peto
      peto
      2009-02-10

      (Sorry, had to wait a little while to be able to test with a book I have; this is also a different scanner...)

      http://img8.imageshack.us/img8/1188/02adjustedag1.jpg

      (LCMS, profile applied in PS, then contrast adjusted.)

      The white has a terrible magenta tint that the profile isn't removing.  If there's any color to the actual paper, I can't see it.  I can remove it manually by levelling up green, which helps a lot but throws the profile out the window (http://img201.imageshack.us/img201/7482/03whitehl7.jpg).

      What am I seeing?  What can be done about it?

      (unadjusted: http://img14.imageshack.us/img14/3688/01rawcd2.png; target http://img17.imageshack.us/img17/6685/targetsmallig2.png\)

       
      • It looks like the media white point of the profile does not match the color of unprinted paper of your book, thus the color of unprinted paper will not map to monitor white, if the profile gets applied with relative colorimetric or perceptual intent.

        You can choose your desired media white point here:

        In "Set Profile Generation Parameters" -> "White Point Handling" -> "Location" select the choice "RGB" and enter the RGB color of the unprinted paper of the book you want to scan (in the given example approx. 246,243,244). Leave "Luminance Scaling" at the default "Do not scale".

        Note, you'll need to build a separate profile (with a different media white point) for each kind of medium you want to scan (unless the paper colors of the media happen to match exactly).

        Regards,
        Gerhard

         
    • peto
      peto
      2009-02-11

      Thanks.

      Is that new to CVS?  I havn't tried CVS yet since building it in Windows is a pain.  I don't mind spending a bit more time trying to get it to build--just checking before I do ...

      I get around 251,245,250--where is 246,243,244 from?

      Does this do something more intelligent than merely using PS levels and selecting the paper as the white point, or does it just effectively bake that same transform into the profile?

       
      • > Is that new to CVS?

        This function is not new, but already quite some time in CVS...
        [not all the choices offered in the GUI are fully implemented yet, though]

        > I get around 251,245,250--where is 246,243,244 from?

        Sorry, my fault, my numbers can't be correct.

        > Does this do something more intelligent than merely using PS levels
        > and selecting the paper as the white point, or does it just effectively
        > bake that same transform into the profile?

        More or less the latter. It's not exatly the same transformation [the WP scaling built into the profile is not done in RGB space], but likely you won't notice much difference.

         
    • peto
      peto
      2009-02-12

      (By the way, rescanned with color balance in vuescan set to "none" instead of "neutral", since some of the light patches were clipping.)

      > This function is not new, but already quite some time in CVS...

      Just wanted to know if it was something I had to build CVS for.  I ultimately did get it to compile (an ... exercise :), and I might post some suggestions to make this easier, if anyone cares...

      > More or less the latter. It's not exatly the same transformation [the WP scaling built into the profile is not done in RGB space], but likely you won't notice much difference.

      I ask because it's a much smoother workflow to scan and adjust everything with the same profile, and then to adjust the whitepoint along with contrast in postprocessing, rather than to have to scan a test page, measure the whitepoint, create a new ICC profile from it, etc.

      I'm still having subjective doubts, comparing the results with someone else's levelling [1].  Basically, things are too bright--the shadows are lighter, the boots are too shiny.  The blue pattern in the background is so close to white it's hard to see the pattern.  I'm not sure how to evaluate this more precisely--just comparing the two results and the original page, these things subjectivally look closer than my current lprof results.

      I tried the "assumed PCS viewing conditions" settings, but that didn't seem to make a difference (maybe not working in CVS right now?).  Playing with levels (monochromatically) doesn't really do it--if I bring it down enough to make the background pattern similarly visible, the rest of the image is maimed.  I don't think it's my monitor calibration--which is not calibrated at all, just a 2405FPW in sRGB, but the differences are similar in all of the monitor presets.

      I'm not expecting an emittive RGB source to look exactly like a reflective CMYK source, of course.  But it seems I should be able to do better, and I'm not sure what to try next...

      [1]
      current lprof: http://img12.imageshack.us/img12/112/test2rp7.png (adopt device whitepoint, auto smoothness, whitepoint adjusted in photoshop, darks levelled up to fix contrast)
      comparison: http://img183.imageshack.us/img183/3770/testwm1.jpg (I don't know what the procedure was for this.)

       
      • > I tried the "assumed PCS viewing conditions" settings, but that
        > didn't seem to make a difference (maybe not working in CVS
        > right now?).

        Try for instance

        Viewing conditions for the captured medium:
        ISO-3664 P1 Critical Print Evaluation Environment

        in conjunction with

        PCS viewing conditions:
        sRGB monitor in darkend work environment

        The corresponding CIECAM02 transformation will be incorporated into the perceptual intent tables of the profile, i.e. you need to apply the resulting profile with _perceptual_ intent (not colorimetric) in order to see the effect.

        [Actually this setting is cheating, since the PCS is actually defined as a virtual reflective print medium. But setting the PCS viewing conditions to sRGB makes still sense, if the scanner profile is supposed to be combined with e.g. an sRGB or AdobeRGB destination profile, which is a matrix profile and therefore colorimetric only, not doing a perceptual transformation to monitor viewing conditions by itself]

        > I'm still having subjective doubts, comparing the results with
        > someone else's levelling [1]. Basically, things are too bright--
        > the shadows are lighter, the boots are too shiny. The blue
        > pattern in the background is so close to white it's hard to see
        > the pattern. I'm not sure how to evaluate this more precisely--
        > just comparing the two results and the original page, these
        > things subjectivally look closer than my current lprof results.

        You can't evaluate this precisely, but only subjectively, since this kind of adjustment [eventually this is a kind of gamut mapping] is no longer purely scientific, but already a matter of art.

        Personally I find that moving the black point in the levels dialog in a gamma-adjusted RGB working space (like e.g. sRGB or Adobe RGB) tends to yield a bit strange results [at least from a colorimetric point of view], so I'd rather tend to do a black point scaling operation either in a gamma 1.0 RGB working space [which will have an effect similar to Adobe's BPC], or I'd do it in a perceptual color space like CIELAB, by moving the blackpoint along the L* axis.

        Regards,
        Gerhard