Excellent tests Sony A350

Discussion in 'Sony' started by Focus, Jun 8, 2008.

  1. As far as I know, no DSLR does binning for good reasons ---
    among them that it's seldom needed, that bayer pattern filters
    don't help, that turning your shiny new 8, 10 or 12MPix-camera into
    - a 4/5/6 Mpix (1x2 binning) camera with non-square pixels
    - a 2/2,5/3 Mpix (2x2 binning) camera
    - a 0.89/1.1/1.3 Mpix (3x3 binning) camera
    is not appreciated by more than 99.9% of the customers!

    Especially as:
    - DSLRs are usually quite noise free and a signal-noise increase
    of 1.4x/2x/3x is not worth the hefty resolution price for
    documentary or artistic photography. (Science and engineering
    often have vastly different needs and restrictions)
    - DSLRs rarely have to cope with so low photon counts and usually
    have quite excellent read noise compared to their sensor wells
    that binning offers nearly nothing over downscaling.
    - Noise reduction based on dark frame substraction works better
    when individual pixels can be targeted as hot or dark and
    interpolated, instead of having a 1/4 hot or 1/4 dark pixel
    (after 2x2 binning) --- and long exposures with dark frame
    substraction is something DSLRs do quite well, thank you.

    Wolfgang Weisselberg, Jun 13, 2008
    1. Advertisements

  2. So you assume that your smaller pixels have less read noise than
    the larger ones ...
    I keep ignoring the fact that you believe you are right.
    Obviously that's a sin.

    Do tell, how do you get better resolution out of a perfect pinhole
    camera by increasing the pixel count beyond reasonable limits?

    Wolfgang Weisselberg, Jun 13, 2008
    1. Advertisements

  3. Avoiding the temptation to ask why binning was even brought up, if no
    DSLRs do it...

    On some small sensor cameras, a reduced-resolution mode is offered at the
    highest ISO levels. This is done by combining pixels, but does anyone
    know whether that combination is true on-sensor binning, or just
    off-sensor pixel addition (resampling to a lower resolution, if you like)?

    David J Taylor, Jun 13, 2008
  4. What is overlooked in the alleged benefits of binning is that binning
    does *not* reduce visible noise, if you maintain the same subject viewing
    size. Binning only reduces noise when it results in a smaller viewed
    image, and that, theoretically, could be done by simply having a finer
    display medium, and maintaining the original pixels. I believe that when
    display devices become much finer than they are now, the folly of big
    pixels and binning will be much more readily apparent; current display
    technologies (and fast dowsizing algorithms) are biased against high-MP

    Hardware binning, on the sensor, may reduce read noise, but probably not
    as well as some of the more optimistic projections; there is more read
    noise incurred after the binning process, for starters.
    Actually, DSLRs are more affected by read noise than P&S cameras are.
    P&S cameras' range of read noise, relative to signal, overlaps DSLRs by a
    good margin, with many P&S cameras having less than some DSLRs. Measured
    read noises for P&S cameras can be deceptive when compared to DSLRs. P&S
    makers have been able to develop extremely efficient microlenses, and the
    native base ISOs of P&S cameras is generally ISOs 160 to 200. The
    manufacturers label them as 80 and 100, possibly for marketing reasons -
    "What? No ISO 100? This camera forces you to have noisy pictures at ISO
    200!". So, when you see a figure of 3 or 4 electrons or 2.5 to 4 ADU of
    read noise in a P&S camera, what you are not being told is that P&S
    cameras have twice as strong a signal in photons/electrons or ADUs as
    most DSLRs. My FZ50 has a read noise of 2.7 ADU at "ISO 100". When you
    meter a white wall externally, or with the meter in the camera, the
    average value of the RAW data is about 700 ADU. In all of my Canon
    DSLRs, and in most DSLRs in general, it is about 350 to 400 ADU. So, a
    P&S with 2.7 ADU of read noise actually is about 1.35 ADU in 12-bit DSLR-
    speak; that is better than all DSLRs except the Canon mk2 and 3 1-series,
    the Fuji S series, and the D3. The noisiest P&S I've tested has a read
    noise of about 4.0 at base ISO ("80"/160), which, using the true (or
    DSLR-normalized)sensitivity, is really about 2.5 ADU, which is better
    than most non-Canon DSLRs.

    So, read noise has about the same isolated effect in a P&S as it does in
    a better DSLR; when combined with shot noise, in quadrature, the read
    noise is less significant in the P&S than it is in the DSLR.
    John P Sheehy, Jun 14, 2008
  5. None of the cameras that do it, do it in RAW mode, even if they have RAW
    mode, AFAIK. I'm sure the technology is currently too expensive for P&S
    cameras, anyway. Besides, a P&S camera's biggest noise problem is shot
    noise, not read noise, and read noise is the only thing that hardware
    binning can potentially improve, at the image level.
    John P Sheehy, Jun 14, 2008
  6. The frequency spectrum of noise, relative to pixel spacing, is similar
    for all pixel spacings. Therefore, if you oversample so that it takes 8
    pixels to get from white to black, most of the noise will be at a much
    higher frequency than the highest frequency of the subject detail;
    "image-smearing NR" will not smear much of anything of value, nor will
    you confuse noise with detail in large murals or fine prints that can
    show every original pixel.
    That is complete nonsense.

    While image-level shot noise stays roughly the same, regardless of pixel
    density, image-level read noise generally *DECREASES* with higher pixel
    density. There is a technological barrier to reading any pixel at base
    ISO with a maxsignal-to-readnoise level greater than about 3000:1. The
    Canon mk2 and mk3 cameras, and the D3 all have approximately this ratio.
    The *only* DSLR to break this barrier, AFAIK, is the Pentax K10D, with a
    ratio of 4500:1 at ISO 100 (if you get a good copy that doesn't have the
    staggered dashes in the deep shadows, it has the greatest DR at the pixel
    level, and its only DSLR competitor at the image level is the 1Ds3).

    Now, take my Panasonic FZ50. It has a base ISO read noise of 2.77 ADU
    (lower than most DSLRs!). The Canon 1D2 has a read noise of about 1.27
    ADU (actually, a little bit higher, relative to saturation, or DR-wise,
    since saturation is far below 4095 in the 1 series). 16.5 FZ50 pixels
    populate the same area as one 1D2 pixel, so 16.5 FZ50 pixels combined
    give a read noise of 2.7 * 16.5^0.5 = 10.97 ADU, with a saturation of
    4058 * 16.5 = 66,957 ADU. that's a ratio of 66,957:10.97, or 6104:1, a
    stop more DR than the 1D2 pixel (3600/1.27, or 2835:1). This is also
    done at a higher true ISO, as the FZ50 is under-rated by a full stop
    while the 1D2 is under-rated by only 1/3 stop, so the read noise relative
    to absolute signal at base ISO for the virtual FZ50 super-pixel is
    actually about 1.67 stops lower than the 1D2 pixel.
    John P Sheehy, Jun 14, 2008
  7. Focus

    Mr.T Guest

    Another black pot.
    Yes, pity it's necessary for people who can't, (or refuse to) read what's
    been written.

    Mr.T, Jun 14, 2008
  8. Thanks, John. I thought that was the case, and that "on-sensor binning"
    as described was something outside the current use of digital camera

    David J Taylor, Jun 14, 2008
  9. Take a shot of an evenly dark surface, with a sensor with a high
    per-pixel noise. Display it at a size and where you can still
    distinguish every pixel.

    Repeat, but use 10x10 downscaling (not even binning) and
    interpolate the missing pixels.

    You'll see a lot more noise in the first image.

    => Your claim, as you presented it, is not true.

    Binning != downscaling.

    Use a sensor where read noise dominates (very low photon counts
    do that to any sensor). What happens when you apply a read noise
    of 10 to a photon count of 5?

    What happens when you use 4x4-binning?

    What happens if you downscale the image instead of binning?

    5 photons have SQRT(5) noise.
    + 10 read noise gives SQRT(SQRT(5)^2 + 10^2) == SQRT (5+10^2) ==
    10,25 noise.
    s/n ratio: 0,45:1

    5 photons and 4x4-binning: 80 photons, SQRT(5*4*4) == SQRT(80)
    +10 read noise: SQRT(80 + 10^2) == 13,42 noise.
    s/n ratio: 5,96:1

    5 photons AND their read noise added up:
    80 photons, SQRT((5 + 10^2)*4*4) == SQRT (1680) == 40.99 noise.
    s/n ratio: 1,95:1

    Even assuming you had a point when it comes to 'downscaling per
    algorithm' not being better than 'downscaling by noise integration
    of the eye', binning is not downscaling.

    Tell the scientists that use it every day that the don't know
    what they do. They'll be very glad to hear from you that
    they've been wrong all along.
    Read noise increases with the size of the charge read --- which
    is all what binning does to the reading? Most interesting!
    Please provide proof!
    Pleae show proof.
    No, you ranting below isn't proof.
    .... so?
    If the United States of America has the same or more idiots than
    some village in outer Elbonia, consisting entirely of idiots
    .... does that mean that the USA is affected by idiots the same
    or more than said village?

    I dare say "NO", you claim yes.
    Irrelevant, microlenses are only combatting a less than 100%
    fill factor or --- seen from another angle --- reduce the
    need to amplify the signal. Doesn't do anything to read
    noise, only changes the base ISO.

    Anyway, DSLRs have the same extremely efficient microlenses.
    I see.
    - How comes that my P&S cameras meter and expose just as expected
    on ISO 50, 100, 200, 400, etc?
    - How comes the metering and exposure settings match that of DSLRs and
    light meters within 0.5 stops or better?
    - How comes the noise gets progressively worse with
    increasing ISO settings?
    - How comes that the exposure time is indeed what's metered
    and shown in the EXIF? (Yes, I tested, CRT monitors are
    quite good for that --- make sure you use portrait mode for
    teh DSLR because of the rolling shutter syndrom), the pattern
    changes distinctly between one stop and the next)
    - How comes the aperture is indeed what's metered and shown
    in the EXIF? (Yes, I tested: same focal length (by mm, w/o
    crop factor), same aperture, same distance ... same DOF!)
    If you meter a white wall, you simply find the bias of the
    meter as to what is "middle gray" --- nothing else.

    700 ADU on a scale of 0-2800?
    350-400 ADU on a scale of 0-1500?

    Comparing apples to oranges.
    Even *if* that was true and comparable, it wouldn't mean a
    thing. How large, did you say, were the sensor wells?
    Ah, yes, you believe that there's no difference between 10 and
    10,000,000 electrons when it comes to noise. Which is a nice
    religion, but my belly laugh.

    Wolfgang Weisselberg, Jun 14, 2008
  10. The frequency spectrum of noise is what you expect of white
    noise, yes.
    So basically, you add very wide band (i.e. white) noise and a high
    pass filter. Interesting idea --- it will surely "remove" (blur,
    or if you want, move to lower frequencies) "all" the high frequency
    noise, but how should it help with the low frequency noise,
    which you *also* add?

    Sounds to me at the very best (and ignoring some facts) as a
    zero-sum game.
    a) It's easy to contrive an image where detail is ath the
    1-pixel-level and cannot be distinguished from noise
    b) unless you have a noise-detail-discriminator as advanced and
    capable as our brain in it's million years of evolution, your
    point's as good as saying "man can easily fly if given wings
    to beat the air, because birds can do it".
    Just because you don't understand things, they are not nonsense.
    That claim is provably false ... and too rubbery in all other
    cases to be of value. What is "roughly"? 10%? A factor of 3?
    Whatever "image-level read noise" is supposed to be, and how it
    combines, is --- of course --- completely undefined. However,
    the sum of read noise over the image increases as the pixel count
    increases, so your claim is either wrong or worthless as it lacks
    definition --- look at "at night it's colder than outside" and
    see how well defined your claim is.
    Proof of the "technological barrier"?

    D200: 3268:1 at ISO 100.
    1D Mk II: 3190:1 at ISO 100.
    So the barrier is anything but unpassable.
    Please provide a reliable source for your 2.77 ADU read noise
    claim at "base ISO" (ISO 80, I presume).

    ISO 100:
    1D MkII: 1.27 ADU
    20D: 2.04 ADU
    350D: 2.12 ADU
    D200: 1.25 ADU

    ISO 200:
    D3 1.07 ADU
    D300 1.20 ADU
    D50 1.80 ADU

    All as per
    and 12 bit ADUs --- just like your FZ50. Yes, that's all
    DSLRs where the data was easily computable. None's as bad as
    your FZ50.

    Why? Because DSLRs collect *many* times more photons, so a few
    more electrons read noise comes out as _much_ less. When it takes
    ¨10-16 electrons to produce a single DN, even 10 electrons read
    noise isn't doing much ...
    It records 53,000 electrons at a gain of 13.02 at ISO 100.
    It records 3,310 electrons at a gain of 0.81 at ISO 1600.
    53,000 / 13.02 = 4071. That's 4071 of 4096 possible values.
    3,310 0.81 = 4086. That's 4086 of 4096 possible values.
    Yes, except for the "LO" iso setting, it's similar for the rest.

    In other words: you don't know --- again --- what you talk
    about --- and if your claim of "4058" further down can be
    believed, the FZ50 is WORSE than the 1D Mk II!
    The FZ50 has 3648 x 2736 pixels on a 7.176 x 5.319 mm sensor.
    The 1D MkII has 3504 x 2336 on a 28.7 x 19.1 mm sensor.
    (data from dpreview.com)
    FZ50 pixel pitch: ~1.93µm.
    1D MkII pixel pitch ~8,22µm.

    (8,22 / 1,93)^2 = 18.14 --- NOT 16.5.

    So it takes ~18.14 pixels on the FZ50 to cover the same area
    as one 1D MkII pixel.
    2.77 ADU * SQRT(18.14) / 18.14 / 0.513 = 1.26 ADU.
    Surprise --- the read noise is within rounding error of the
    1D MkII, assuming your claim of 2.77 ADU is correct.

    The 0.513 is the conversion factor between ISO 80 (what the
    FZ50's base ISO is) and ISO 125 (What the 1D Mk II uses when
    it's set to the base ISO of 100 --- see dpreview.com!)

    So now you have a 856 x 642 pixel camera --- not even 0.55MPix,
    with identical read noise. In other words, you have turned
    a photon noise limited FZ50 into a photon noise limited 0.55M
    webcam replacement. Congratulations!

    Of course, the 1D MkII can scale down to 856 x 642 pixel
    quite easily and get 0.33 ADU read noise.
    Proof not offered, conspiracy theory detected, claim
    rejected as "completely off the rocker".

    Wolfgang Weisselberg, Jun 15, 2008
  11. A CCD is more prone to thermal noise, so harder to keep quiet in live
    view mode when it is continuously operated. But the A350 doesn't use
    the image sensor for live view, it uses a small accessory screen for
    that. So they may have chosen the CCD because of its nicer noise

    My impression is that Sony doesn't usually compromise on engineering
    quality when trying to keep costs down, preferring to omit features
    I can't remember what TV shots of Japanese papperazi thronging around
    a celebrity in Tokyo I spotted recently, but one thing which struck me
    was that a few of them were using big MF cameras. It wouldn't surprise
    me if there's a lot more enthusiasm for big high pixel count sensors
    amonst professional Oriental photographers than there is in the
    West. And as it happens, they happen to be the people making the
    Chris Malcolm, Jun 18, 2008
    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.