Lack of deep shadow detail in digital images

Discussion in 'Darkroom Developing and Printing' started by Nicholas O. Lindan, Jan 6, 2006.

  1. As part of a post on dpi Vs. pix/inch in printers, I made the following
    calculations:


    If you use 8 bits in digital imaging and process for a 2.0 OD
    for dead black, the standard in chemical photography, the
    amount of detail in the deep shadows is really limited:

    pixel value OD Zone Shades/Zone change
    255 2.0 0
    254 1.86 I 1
    251 1.59 II 3
    245 1.31 III 6 - detailed black
    232 1.00 IV 13 - dark grey
    1 0.0017
    0 0.00

    I'm not a big digital fan, but the times I have tried scan/print
    I have been frustrated by the blocked up shadows. I think I see
    why. Rich deep velvety blacks are not in the cards.
     
    Nicholas O. Lindan, Jan 6, 2006
    #1
    1. Advertisements

  2. Nicholas O. Lindan

    Lorem Ipsum Guest

    Nicholas, you are correct, of course. Someone will dive in and tell you to
    make multiple layer exposures. Screw that.

    On the other hand, you might want to experiment with Tiffen's 'contrast''
    filters. They lower contrast. 'course, with the impoverished real-estate of
    photo-cells, you may as well use that other stuff, what do they call it,
    FILM! Yeah. That's it. (I had to Google it out.)

    ;)
     
    Lorem Ipsum, Jan 6, 2006
    #2
    1. Advertisements

  3. Nicholas O. Lindan

    Scott W Guest

    I am having a hard time following where you got your numbers. I am
    assuming that if you are setting the OD of 2 to a pixel value of 255
    you are dealing with a scanned negative. If so then by the time you
    are to an OD of 1 this should be represent about 10% to 1% of the light
    needed for an OD of 2, depending on the film and the development.
    Let's take the high contrast case of 10%, this would mean that an OD
    of 1 should map to a pixel value of close to 90, I don't see how it
    could possibly map to 232. 232 is not anywhere near a dark grey.

    Scott
     
    Scott W, Jan 6, 2006
    #3
  4. 8 bits is enough for moving images; e.g., movies. When I was doing
    Picturephone research, 8 bits were barely adequate for moving images; 7 bits
    were objectionable. And for still images, you need a lot more because the
    eye-brain system gets more time to evaluate what it is seeing, and is more
    critical.

    For still images, more are required unless you use some kind of compression.
    In computer-speak, if you are going with 8 bits, you better use a floating
    point representation or other forms of lossless compression. It is simpler
    to go to 12 to 16 bits. Some very expensive scanners will give 16 bits or
    more (e.g., Howtek).
     
    Jean-David Beyer, Jan 7, 2006
    #4
  5. Nicholas O. Lindan

    nailer Guest

    what about gamma? have you tried to manipulate the shape of the
    function? it doesn't have to be linear.


    #As part of a post on dpi Vs. pix/inch in printers, I made the
    following
    #calculations:
    #
    #
    #If you use 8 bits in digital imaging and process for a 2.0 OD
    #for dead black, the standard in chemical photography, the
    #amount of detail in the deep shadows is really limited:
    #
    # pixel value OD Zone Shades/Zone change
    # 255 2.0 0
    # 254 1.86 I 1
    # 251 1.59 II 3
    # 245 1.31 III 6 - detailed black
    # 232 1.00 IV 13 - dark grey
    # 1 0.0017
    # 0 0.00
    #
    #I'm not a big digital fan, but the times I have tried scan/print
    #I have been frustrated by the blocked up shadows. I think I see
    #why. Rich deep velvety blacks are not in the cards.
     
    nailer, Jan 7, 2006
    #5
  6. Nicholas O. Lindan

    W (winhag) Guest

    If you are talking about digital cams, I suggest you start shooting and
    stop calculating.
    My practical experience has been that overall image quality for a given
    format is clearly
    superior with digital. If you are talking about scans, you can always
    go 16 bit.
    However, I think for consumer scanners it is the CCD that is probably
    the limiting factor
    more than the number of bits.
     
    W (winhag), Jan 7, 2006
    #6
  7. Nicholas O. Lindan

    Tom Phillips Guest

    Format? There are various types of digital scans
    and scanning sensors. Most common is bayer pattern
    1-shot. These don't offer anything but one type of
    "format." Trilinear sensors might be said to offer
    a different format, but also require long exposures
    due to the focal plane scans. Format refers to camera
    size, not sensor type...

    If you are talking about scans, you can always
    Nicholas is talking output. You're talking input.
    Not the same.
     
    Tom Phillips, Jan 7, 2006
    #7
  8. Nicholas O. Lindan

    Tom Phillips Guest

    Yes but the issue as identified here is the resulting
    inkjet output, not the scan...
     
    Tom Phillips, Jan 7, 2006
    #8
  9. Jean-David Beyer wrote:
    []
    Was that 8 bits linear or 8-bits with a gamma correction of around 2.2?
    My tests have shown that, with gamma-corrected images it's quite hard to
    tell between 7-bit and 8-bit quantisation except with special test
    patterns.
    8-bits, gamma-corrected, already has a greater dynamic range than most
    12-bit camera sensors, but you loose out if you need to restore shadow
    detail.

    David
     
    David J Taylor, Jan 7, 2006
    #9
  10. I do not remember. This was in the late 1960s and early 1970s. If the TV
    cameras were gamma corrected (they used vidicon tubes, IIRC; they were not
    the studio image-orthocon tubes), so were we, and otherwise not. We used a
    linear A to D convertor, and at television speeds the best you could get for
    $10,000 was 8-bits linear. At the output of our system, we used 12-bit D to
    A convertors, since they were simple. I am pretty sure they were linesr, but
    we used them to drive standard monitors of the day.
    I would say that gamma correction is just an approach to imitating floating
    point representation of the data, where the counts were close together near
    zero illumination, and far apart far from zero.
     
    Jean-David Beyer, Jan 7, 2006
    #10
  11. I remain very surprised that you found 7-bit data "objectionable". It
    flies in the face of other research done at the time, even on PAL colour
    systems. The one area where 8-bits might not be adequate (as far as I
    know) is in RGB colour when you are looking at fine colour differences.
    Was your work monochrome or colour?

    Yes, gamma correction does allow a greater dynamic range at the expense of
    accuracy in shadow detail. Fine for purely display purposes, but lacking
    when significant processing is required. Even 12-bits seems to be not
    quite adequate for the best of today's sensors.

    David
     
    David J Taylor, Jan 7, 2006
    #11
  12. We were doing monochrome initially. The problem with 7-bits is that
    contouring is to easily seen. It is less obvious in moving images than in
    still ones, but still objectionable by A/B testing where the observers
    evaluate which they prefer. The observers (and we kept running out of those,
    since we wanted "untrained observers" and by participating in the studies,
    they became "trained) preferred the 8-bit to the 7 bit with great
    statistical significance.

    When we went to color, we used 16-bits altogether. IIRC, we used 7 bits for
    green (most like luminance), 5 bits for red, and 3 bits for blue (eye not
    too sensitive to that). That was considered acceptable by the observers.
    I guess it depends on how you do it. It seems to me you could use more bits
    in the shadow areas and less in the highlights to get better shadow detail.
    Sort of like we did for audio stuff. With not enough bits near zero, the
    initial parts of words, which are softer (unless they are percussive) get
    clipped and this is very objectionable. So what gamma correction you apply,
    by whatever measure you choose, depends on what part of the dynamic range is
    most important, including signal to noise ratio.

    If wet process photography is any guide, most people like the middle of the
    range (e.g., Zone III to Zone VIII) to be well rendered and take what they
    get with the rest. Now "fine-art" photographers can control this better than
    a mass photofinisher, but they are not typical.

    And adding color, especially red ;-) , makes things more acceptable to the
    naive user.
     
    Jean-David Beyer, Jan 7, 2006
    #12
  13. I realise it's probably 30 years too late (<G>), but I would imagine that
    adding a small amount of dither would have obscured the contouring on
    7-bit monochrome.

    Noted on the 16-bit colour - I've played a little with this on paletted
    displays, but as soon as you start getting into colour it's a whole lot
    more subjective. I can see how your RGB split is reasonable.

    I had a colleague who built an audio processor using a log representation
    of the signal - great dynamic range and easy multiply (add the logs), but
    signal addition was a pain! Of course, for 8-bit audio there was (is?)
    u-law and things like that, sort of the equivalent of gamma-adjusted image
    data. Interesting to see now the change from 16-bit to 24-bit and higher
    accuracy linear audio processing. I wouldn't be surprised to see 16-bits
    become the standard for images in a few years time, perhaps just in DSLRs
    and not the consumer market, though.

    Cheers,
    David
     
    David J Taylor, Jan 7, 2006
    #13
  14. Yet many digital cameras cover an 11-12 stop dynamic range ...

    Bart
     
    Bart van der Wolf, Jan 7, 2006
    #14
  15. Dynamic range and number of bits in the data have nothing to do
    with each other:

    Count Intensity
    ===== =========
    255 Surface of the sun
    -
    0 Blackbody at absolute zero

    -OR-

    Count Intensity
    ===== =========
    255 Pale grey card on a cloudy day
    -
    0 An ever so lighter shade of pale grey


    * * *

    Maybe this is the wrong group for this discussion.
     
    Nicholas O. Lindan, Jan 7, 2006
    #15
  16. Nicholas O. Lindan

    Frank Pittel Guest

    : As part of a post on dpi Vs. pix/inch in printers, I made the following
    : calculations:


    : If you use 8 bits in digital imaging and process for a 2.0 OD
    : for dead black, the standard in chemical photography, the
    : amount of detail in the deep shadows is really limited:

    : pixel value OD Zone Shades/Zone change
    : 255 2.0 0
    : 254 1.86 I 1
    : 251 1.59 II 3
    : 245 1.31 III 6 - detailed black
    : 232 1.00 IV 13 - dark grey
    : 1 0.0017
    : 0 0.00

    : I'm not a big digital fan, but the times I have tried scan/print
    : I have been frustrated by the blocked up shadows. I think I see
    : why. Rich deep velvety blacks are not in the cards.

    To me it doesn't matter much how I get the image onto paper. To me digital
    prints just don't have the look of silver gelatin prints. Once it is
    though I'm going to pack up my enlarger and go to digital printing!!
    However as you point out that day is a long way away for more reasons then
    the ink!
    --
     
    Frank Pittel, Jan 8, 2006
    #16
  17. If you were digitizing the output of a video camera, and using it to
    directly provide video to a CRT, then you *were* working in a
    nonlinearly encoded space. More precisely, the encoding is close to a
    power function:

    Vout = normalized_light_intensity ^ 0.45
    There are a whole bunch of representations that you might imagine using.
    Floating point itself can be seen an easily-implementable approximation to
    storing the logarithms of numbers. But it turns out that a power
    function with an exponent of somewhere around 1/2 to 1/3 matches the
    eye's "lightness" response pretty well, and thus makes good use of the
    available bits. If you also pick the exponent to be about the inverse
    of the CRT voltage-to-current transfer curve, then you end up with an
    encoding scheme that (a) doesn't need compensation for CRT nonlinearity,
    and (b) makes perceptually better use of the available bits than almost
    any other method.

    I've done some of my own tests, and 8-bit "gamma encoded" data seldom
    shows quantization artifacts even with noiseless (computer rendered)
    source images, while using a logarithmic coding with a reasonable
    intensity range (e.g 200:1) shows banding until you have at least 9
    bits/pixel. I was surprised by this result at first, but it turns out
    that the eye's sensitivity to banding depends on brightness. In bright
    areas, you can see less than 1% change in brightness across an edge -
    and 8-bit gamma encoding provides this. In dark areas in the same
    image, changes of a few percent are undetectable - and this is all gamma
    encoding provides. Logarithmic encoding gives the same relative step
    sizes over the whole brightness range, which sounds great, but 8-bit log
    encoding isn't good enough for bright areas. (And 8-bit linear is
    totally hopeless in the shadows).

    Dave
     
    Dave Martindale, Jan 8, 2006
    #17
  18. I cannot figure out how you generated this table. Could you provide
    more info? Are you taking 255 to mean white (the usual convention in
    digital photography) or black (if you're talking about ink coverage in
    printing)? And is the OD value density on the print?
    Many printing processes are not capable of deep blacks, particularly if
    they don't print on glossy papers. But the 8 bit coding is not the
    problem, because pixel values don't map to densities in the way you
    calculated above. If I have time later, I'll generate something more
    plausible.

    Dave
     
    Dave Martindale, Jan 8, 2006
    #18
  19. SNIP

    Indeed:
    http://www.xs4all.nl/~bvdwolf/main/foto/Imatest/1dsm2/1436_Step_2.png
    Shows almost 11 stops, with lens flare.

    I'm close to completing my new test set-up to avoid flare and to make
    it suited for any lightsource (not only Tungsten like in the example,
    which thus results in higher noise in the blue and green channel).
    Maybe...

    Bart
     
    Bart van der Wolf, Jan 9, 2006
    #19
  20. Lots of this has been known for many many decades. The work of Helson and
    Judd at the National Bureau of Standards, probably before WW-II, dealt,
    among other things, with these very aspects of human visual perception.
     
    Jean-David Beyer, Jan 9, 2006
    #20
    1. Advertisements

Ask a Question

Want to reply to this thread or ask your own question?

You'll need to choose a username for the site, which only take a couple of moments (here). After that, you can post your question and our members will help you out.