Bayer aliasing on the 5D

Discussion in 'Digital Photography' started by Alfred Molon, Jan 28, 2006.

  1. I think you are referring to this closing of a thread exchange:
    The tiny effect (in printed(!) output, we're talking about a centre
    blur diameter of approx. 1 to 2 mm easily resolved with a
    loupe), can be made bigger by using more cycles per
    circumference. Also a zone-plate structure will offer more high
    frequency signal. However, a large benefit of a star chart is
    that its evaluation in absolute terms is insensitive to (reasonable)
    variations in shooting distance.

    The (luminance) Nyquist frequency of a sampled capture of e.g.
    a 60 cycle star is always 38.2 pixels in diameter (which is
    already significant on most monitors at a 100% zoom setting).
    Any 'detail' within that circle (especially hyperbolic deviation
    from the underlying structure) is aliasing. The hyperbolic
    deviations (due to angled arrangement of the cycles) also avoid
    Phil Askey's (DPreview) fallacy that Foveon's non-AA-filtered
    aliasing is useful because it resembles real detail.

    It is also trivial to make a star chart with e.g. 120 cycles available
    for download (Epson and/or Canon/HP native resolution). I use such
    a target myself to judge the aliasing effects of Digicams (and Raw
    converters !).

    The only drawback is that it requires a bit more accurate
    (photo-grade glossy paper, preferably with a good paper profile)
    printing because gamma adjusted gradients become more critical.

    This is an example of an Adobe Camera Raw 3.3 Raw conversion
    (no color noise reduction, no sharpening) of a TS-E 90mm f/2.8 lens
    I tested on an EOS-1Ds Mark II this week:
    <http://www.xs4all.nl/~bvdwolf/main/foto/Imatest/1dsm2/1953crop.png>
    If the Nyquist limit is marked
    <http://www.xs4all.nl/~bvdwolf/main/foto/Imatest/1dsm2/1953crop+.png>
    it becomes clear that the most obvious artifacts occur at Nyquist or
    smaller (due to lack of other signal), with some false color
    artifacting (due to lower sampling rate by Bayer CFAs) as the
    Nyquist limit is approached.
    Which also requires, as you hinted at, that the dot pitch of the
    TV/monitor as projected on the sensor is significantly smaller than
    the sensor's Nyquist frequency. Otherwise it becomes simple moire
    instead of aliasing.
    Two observations. First, as Canon states, the phase based AF
    system may be 'confused' by regular patterns. So focus
    (confirmation) may be inaccurate when confronted with aliased
    (=my translation) structures.
    Second, the focus spot feedback on the viewscreen is smaller
    than the actual linear/cross sensor. Also there may be a slight
    mis-alignment of the focal-spot indicator and the actual AF-
    sensors. So, if shot at an angle, the actual AF sensor may
    pick-up something different than indicated in the viewscreen.

    To also add a third (over-hyped) observation, there could be a
    slight calibration error between body and lens. That's a serviceable
    issue, although I'd like to stress that it either applies to all
    lenses (thus indicating a body/mount deviation), or differs by
    lens (thus indicating a lens calibration issue). Again, it's
    serviceable, although it might mean some time without a camera
    + lenses. There is also a (half DOF for consumer, and 1/3rd DOF
    for Pro-model) body criterion that needs to be satisfied for
    calibration needs. If focus is within that margin, they will be
    reluctant to 'repair' the issue.

    Bart
     
    Bart van der Wolf, Feb 3, 2006
    #41
    1. Advertisements

  2. SNIP
    You'll find that at Nyquist and higher frequencies, there is still a
    significant level of modulation. Even on the EOS-1Ds Mark II, which
    has a very decent AA-filter, I often find (image center) modulations
    of 15-20% at Nyquist. That is with all ACR converter settings like
    sharpening and tone curves turned down or at neutral. The EOS-5D will
    no-doubt exhibit even higher modulation (all the rest being equal).

    Here's an example of the TS-E 90mm f/2.8, which peaks at wider
    apertures than I expected:
    <http://www.xs4all.nl/~bvdwolf/main/foto/Imatest/1dsm2/1953_ACR33_YL9_cpp.png>
    The fact that I rarely see aliasing artifacts in my (e.g.
    architecture) images is without doubt helped by the fact that DOF (I
    calculate with a CoC of 2x sensel pitch) is actually quite shallow, or
    at smaller apertures resolution is diffraction limited. The 'defocus'
    will reduce the modulation quite effectively (at wider apertures).
    This is the same evaluation as above, only a very minor defocus was
    used:
    <http://www.xs4all.nl/~bvdwolf/main/foto/Imatest/1dsm2/1975_ACR33_YL9_cpp.png>

    Interestingly, although this defocus reduces the modulation of the
    aliasing artifacts, it also triggers the ACR converter into generating
    other artifacts at lower frequencies:
    <http://www.xs4all.nl/~bvdwolf/main/foto/Imatest/1dsm2/1975_defocus.png>

    Finally, although slightly OT, I'd like to share my findings with the
    MTF50 performance of the above mentioned TS-E lens over the full range
    of apertures:
    <http://www.xs4all.nl/~bvdwolf/main/foto/Imatest/1dsm2/(De-)focus_vs_Diffraction.png>
    This shows that this lens 'peaks', unlike most of my other lenses, at
    relatively wide apertures. It also shows that, obviously, the effect
    of a given amount of defocus is diminishing as the apertures get
    smaller and diffraction takes over.

    And to get back on topic, as diffraction increases, the modulation at
    and beyond Nyquist also reduces:
    <http://www.xs4all.nl/~bvdwolf/main/foto/Imatest/1dsm2/Nyquist_Modulation.png>

    Bart
     
    Bart van der Wolf, Feb 3, 2006
    #42
    1. Advertisements

  3. SNIP
    Actually, what you're seeing is aliasing, because the sampling density
    for color is much lower than for luminance. Added to that, there may
    also be some demosaicing failure involved, but Raw converters are
    becoming pretty good.

    SNIP
    It is also a nice compromise, given current technology, between
    sampling density and dynamic range (well depth). Squeezing in more
    sensels will result in reduced DR because of the smaller sensels. For
    significant advances in resolution (magnification potential), larger
    sensors will be required. It might also mean that Canon starts to
    introduce their full RGB per sensel technology after their upcoming
    22MP models (which may turn out to be a compromise, unless 'binning'
    is part of the proposition to improve DR).

    Bart
     
    Bart van der Wolf, Feb 3, 2006
    #43
  4. Alfred Molon

    W (winhag) Guest

    ...It might also mean that Canon starts to
    Bart,

    Have you heard that Canon has 'full RGB per sensel technology' a la
    Foveon?

    W
     
    W (winhag), Feb 3, 2006
    #44
  5. SNIP
    I've read their 1996 (yes, almost ten years ago) Patent (6,388,709)
    paper "Image sensing apparatus with optical modulation elements having
    transmission characteristics controllable by pixel". It's nothing like
    Foveon, it's more like fast switching electrochromic filter layers and
    repeated readouts. I have no idea how well it would work, especially
    with subject movement.

    Bart
     
    Bart van der Wolf, Feb 3, 2006
    #45
    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.