Bayer aliasing on the 5D

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

  1. SNIP
    That's too arty for daily use.

    Depending on the DOF requirements, either very slight defocus of the
    offending structures, or stopping down beyond the sweet-point (varies
    by lens but that's usually around f/8.0 with 35mm DSLRs), may limit
    resolution just enough to limit the aliasing somewhat.

    The real problem is that you usually can't predict it unless viewed at
    full size after Raw conversion. Going back to reshoot isn't always an
    option.

    In general I'd be careful with shooting architecture (repetitive
    structures), especially at an angle, and fabric.

    Bart
     
    Bart van der Wolf, Jan 30, 2006
    #21
    1. Advertisements

  2. Looks more like corrugated glass roofing to me as it is present on the
    sky seen through that segment. Perhaps glass with wire reinforcing?
     
    Kennedy McEwen, Jan 30, 2006
    #22
    1. Advertisements

  3. I'll meander by there today with a longer lens and see...

    David J. Littleboy
    Tokyo, Japan
     
    David J. Littleboy, Jan 30, 2006
    #23
  4. It is certainly still present at f/8 on all of the following lenses - I
    checked after David pointed me to this image a couple of weeks ago when
    I commented that my first attempt to induce aliasing had failed. I have
    now checked:
    24-105mm f/4L IS USM
    16-35mm f/2.8L USM
    100-400mm f/4.5-5.6L IS USM
    50mm f/1.4 USM

    Quite surprising really as the diffraction limited MTF is only 30% at
    the Nyquist of the sensor and the combination of lens and sensor
    (assuming 100% fill factor at 8.2um) never exceeds 4% beyond Nyquist (at
    1.28x Nyq.).

    The test I did was to create an image of a few hundred 1 pixel black
    lines on a white background and then print this on a standard laser
    printer at 100ppi. I then stuck this on a wall and shot it at focal
    lengths and ranges which would produce aliasing if it was present. For
    example, at 2m range, the Nyquist occurs at about 65mm, and the range
    and focal length can be scaled accordingly to go 20-30% beyond the
    limit. All the tests were made with the chart in the centre of the
    field, working on the assumption that lens MTF would be best there and
    so the problem would show up at its worst. Even at f/11 there is some
    trace of it, but not enough to worry about for real images.
    Even more surprisingly, I think you can predict it. On the tests I
    conducted I noticed aliasing *in the viewfinder* on almost all of the
    shots that produced aliasing on the final result. This was totally
    unexpected because the viewfinder is a quite separate light path from
    the sensor. The only explanation I have is that the precision matte
    screen in not a random ground glass pattern, but a fine regular grille
    or matrix. I first noticed aliasing in the viewfinder while shooting
    some architectural scenes last weekend - and used it to get accurate
    vertical alignment, in the absence of any other alignment guides on the
    standard screen. At the time, it crossed my mind that it might be a
    useful manual focus aid in some cases, but I haven't tried that yet.

    I can't imagine that Canon deliberately selected a screen grating
    similar to the sensor though - it is probably just one of those flukes
    that happen from time to time.

    PS. I just had some 23x33" (A1) prints made of some of the shots I took
    when I first noticed the viewfinder aliasing - quite stunning though
    just on the verge of visible softness at close inspection. I never
    thought 35mm could be this good!
     
    Kennedy McEwen, Jan 31, 2006
    #24
  5. All my lenses will produce Moire, even on the 300D, despite being far more
    plebian than yours: 17-40, Tamron 28-75, Canon 55-200; the only lens we
    share is the 50/1.4. (Actually, I haven't checked the 55-200 for Moire on
    the 300D; but I've seen Moire from it in architectural shots on the 5D.)
    My claim is that the problem occurs in the range between 2/3 Nyquist and
    Nyquist. It's not so much classical Moire as a demosaicing failure. And, of
    course, the wider pixel spacing on the 5D means that it's much more likely
    to occur than on the 20D or other high-res APX-C cameras. Sigh. My bet is
    that the 5D's AA filter might be far closer to adequate were it a 16MP
    camera. Sigh, again.
    An easier test is Bart's star-pattern resolution test chart*, since it will
    have the maximally obnoxious frequencies for every case<g>.

    *: You'll have to ask Bart for the link.
    Whose spacing is an integer multiple of the sensor spacing. Personally, I
    think you're halucinating, but it'd be rude to say so.
    It can't. This isn't film, it's digital. And it's different.

    But seriously, my theory here is that there's a qualitative difference
    between 6 or 8MP and 12MP (or adequately antialiased 16MP). At adequately
    antialiased 16MP, you are capturing enough detail that viewers can come in
    for a closer look and find new stuff that they didn't see at "normal viewing
    distance", so larger prints fly. (At 23x33, the 5D will be 125 ppi and the
    20D will be 100 ppi. I wonder how different that is? (I.e., is my theory
    here completely flaky?))

    It'll be interesting to see what happens if Canon comes out with the rumored
    22MP (3800 x 5700) 1D series camera. That's an 80 lp/mm Nyquist frequency
    with a 55 lp/mm upper limit on practical resolution. Pixels about the same
    size as the 20D. Providing sharp images at the corners will be really really
    hard, even for Zeiss Contax glass, at 50mm and wider. Of course, all the
    test shots will be with telephoto primes and sharp corner to corners.

    So my bet is that adequately antialiased 16MP is the sweet spot for 24x36mm
    digital.

    David J. Littleboy
    Tokyo, Japan
     
    David J. Littleboy, Jan 31, 2006
    #25
  6. Yes, it is also visible below Nyquist, but not as far down as 2/3.
    No need, I already have Bart's tiff file, and I did try it. Whilst it
    certainly produces aliasing, it is not as easy to ascertain how severe
    it is at any particular spatial frequency, because the problem occurs so
    close to Nyquist, where the camera is at its limit of resolution.

    What I intend to do, but haven't got round to it yet, is make a series
    of SFR plots for the lenses at different apertures and field positions
    using the slanted edge method. Whilst not showing aliasing directly,
    that will quantify the magnitude of the response that causes it.
    Do you have details of this? Until I saw this I expected the screen to
    be a fine, but random, pattern - just like most ground glass screens. It
    obviously isn't.
    Certainly not hallucinating David, it is as predictable as the aliasing
    on the final image. Depending on the cladding on the building I
    sometimes found it more accurate to align verticals in the viewfinder
    using this effect by checking with a level. I guess it would go awry in
    Pisa though! ;-)
    It can - it is still 35mm, whether digital or film. ;-)

    Obviously with film there is the added dimension of granular noise which
    increases with spatial frequency, making the image less acceptable than
    an essentially noise free digital source.

    What I meant though, was not relating to the medium as such but the
    magnification associated with printing that format at that size. 125ppi
    is just becoming visible with the naked eye - 250ppi would be as much as
    a typical eye can see - however I expected the optical limitations to
    dominate the visible softness. I am tempted to repeat one or two prints
    after upsampling to 250ppi-300ppi with some USM - adding false detail I
    know, but perception can make as big a difference in this region as real
    information.
     
    Kennedy McEwen, Jan 31, 2006
    #26
  7. Alfred Molon

    W (winhag) Guest

    I have even seen occasional Moire/aliasing on the 20D. I think the
    comments here are pretty accurate reflecting the various vendors
    attitudes towards 'proper' anti-aliasing vs. image 'sharpness' (some
    (all?) Hasselblad digital backs have no anti-aliasing and I understand
    have significant moire/aliasing issues to the point where RAW
    converters have processing to specifically address this). It is
    unfortunate that 'proper' optical A-A filters are not as easy to
    produce as oversampled audio A-A filters are now-adays.
     
    W (winhag), Jan 31, 2006
    #27
  8. Alfred Molon

    Alfred Molon Guest

    Actually since the colour resolution is about half of the sensor
    resolution, you would have to use a very strong AA filter (one with a
    relatively low spatial cutoff frequency), to avoid colour aliasing.

    I also suspect that AA filters are not too steep, which is why some
    manufacturers choose to use too weak ones, to get more sharpness.
     
    Alfred Molon, Jan 31, 2006
    #28
  9. Alfred Molon

    Skip M Guest

    Yet you started this thread by accusing Canon of using an inadequate AA
    filter on the 5D...
     
    Skip M, Jan 31, 2006
    #29
  10. Of course they aren't - they are only sinx/x functions as a consequence
    of the lateral half pixel (assuming the AA filter is manufactured to an
    accurately controlled thickness) shift.

    There isn't any way of making the ideal brick wall filter optically and
    until such times as the CMOS/CCD minimum geometry reduces by an order of
    magnitude or more there isn't any hope of oversampling the image and
    filtering digitally without a massive loss of sensitivity, manifested as
    noise.

    Ideally we would have 0.5um pixels for 0.5um wavelengths, but the pixels
    would have to be deeper and the parasitic track area much reduced to
    prevent it occupying valuable sensitive area. Even then, the resistance
    of the tracks, particularly to central pixels, would result in debiasing
    of the photodiodes and excessive dark current.

    So, realistically, you need X-ray lithography (we are currently only at
    soft UV in production) and room temperature superconductors to get the
    equivalent of audio oversampling. It's not going to happen anytime
    soon. ;-)
     
    Kennedy McEwen, Jan 31, 2006
    #30
  11. Yes, depending on the actual lens' sweetspot, I was thinking about
    stopping down another 1-2 stops, and somewhat 'compensate' for
    diffraction losses by optimized sharpening based on a wider PSF.
    It probably is a type of fresnel lens pattern used to compensate for
    the light fall-off that's common in a simple ground glass. You can
    also notice the viewfinder moire when you focus on a TV screen.

    SNIP
    A fluke, no doubt, but a useful one if it proves to be a decent
    predictor of evil.

    Bart
     
    Bart van der Wolf, Feb 1, 2006
    #31
  12. Alfred Molon

    ASAAR Guest

    A fluke perhaps. But your quote omitted something:
    The fresnel lenses I've seen all have had circular gratings, so if
    that's what was used in the viewfinder how could it have helped
    Kennedy with the vertical alignment of his architectural shots?
     
    ASAAR, Feb 1, 2006
    #32
  13. Most of the screens I have used have a fresnel pattern on them to
    compensate for light fall off, but that is generally circular and
    centred in the field, hence refracting light from the edges and corners
    back to the eyepiece. They don't tend to produce aliasing, at least in
    the centre of the field, because there are no fresnel steps there - and
    they aren't very fine at the edges.

    If the screen was circular then it would not produce symmetric aliasing,
    ie. fine verticals aliased to course verticals etc., but would translate
    fine vertical or horizontal lines to hyperbolics, and fine circular
    structure to straight lines - eg. as with two zone plates overlapped.
    With the Canon screen, fine regular structure is aliased to coarse
    structure in the same axis - just as you would expect from linear
    sampling in horizontal and vertical axes.

    I think it must be a regular x-y structure in the screen texture itself.
    David mentioned in another thread that one of their screens used a fine
    microsprism structure, and that would produce a similar effect, but they
    would have to be very fine microprisms indeed to alias at this level.
     
    Kennedy McEwen, Feb 1, 2006
    #33
  14. I just posted something along the same lines before I read your comment.
    A circular structure on the screen, particularly a fresnel pattern,
    would alias straight lines to hyperbolic lines. I guess the centre of
    the hyperbolic could appear almost straight in the middle of the screen,
    but it would rapidly change size and shape as the original structure
    moved off axis. I didn't notice that happen - but I was more interested
    in correcting the slight error in alignment of my tripod base using this
    unexpected phenomena that trying to examine how it changed across the
    field.

    Nevertheless, I feel instinctively that it has to be an x-y structure in
    the screen texture itself to produce this effect.

    David might be able to through some light on this because from what he
    mentioned in another thread, at least some of Canon's screens use a fine
    microprism structure. I haven't viewed any of the other screens, so I
    can't comment on the differences that they produce.
     
    Kennedy McEwen, Feb 1, 2006
    #34
  15. I just tried that and you are right - it certainly does alias, which
    suggests that the structure itself is much coarser than the sensor pitch
    and what I saw in my tests was probably a harmonic of the aliased
    detail.

    However, if you view the TV screen (or computer screen which is probably
    finer structure) you can see the fresnel lens on the screen as well as
    this aliasing. The fresnel steps appear as discontinuities in the
    pattern.

    So I am convinced now that this is an x-y structure on the focus screen
    - but probably not quite as useful as I had originally expected since it
    will show aliasing when none would be produced by the sensor as well as
    when it would.
     
    Kennedy McEwen, Feb 1, 2006
    #35
  16. I believe that's a one-pixel shift, not a half-pixel shift. One pixel
    is half a wavelength at the Nyquist frequency, which is a 180 degree
    phase shift, so you get the first zero in the filter response at exactly
    the Nyquist frequency, which is what you wanted.

    (Actually, I've heard of people using a smaller shift like 0.8 pixel to
    increase sharpness of in-band detail at a cost of some increased
    aliasing).

    Dave
     
    Dave Martindale, Feb 1, 2006
    #36
  17. I have the grid screen (which is supposed to be pretty much the same as the
    standard screen) installed, and I don't see such an effect with Bart's start
    chart.

    David J. Littleboy
    Tokyo, Japan
     
    David J. Littleboy, Feb 1, 2006
    #37
  18. Hmm, a half pixel shift would give a 30% reduction at Nyquist, a full
    pixel shift would result in a null - but significant loss below Nyquist.
    Who knows what Canon have used on the 5D? I suspect the only way to
    find out is so SFR measurements.
     
    Kennedy McEwen, Feb 1, 2006
    #38
  19. You won't, because the effect only appears right in the centre of the
    pattern - I've already discussed this with Bart on another forum where
    it was suggested as a means of determining printer resolution.

    Bart's chart is good where you can magnify the final output and see
    where the system limits are, but less useful where you have to view the
    result at normal scaling and the effect is tiny. its just the wrong
    tool for this job.

    As Bart suggests, look at a TV screen - or even your computer monitor at
    the appropriate range.

    BTW I have noticed a slight discrepancy between the "in focus" point as
    determined on the screen (maximum alias amplitude) and the autofocus
    point, using the computer screen with a white image on it. I would be
    interested to hear your results of the same thing.
     
    Kennedy McEwen, Feb 2, 2006
    #39
  20. Yes, I get weird Moiré patterns in the viewfinder (at f/2.8) that also show
    up in the captured image (also at f/2.8) from my LCD screen. But this has to
    be a different effect than Moiré in patterns near the Nyquist, since those
    don't happen at f/2.8...

    David J. Littleboy
    Tokyo, Japan
     
    David J. Littleboy, Feb 2, 2006
    #40
    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.