The Joy of Pixel Density

Discussion in 'Digital Photography' started by John P Sheehy, Jul 15, 2008.

  1. John P Sheehy

    Paul Furman Guest

    So an I understand from that, small pixels get muddled data from deeper
    penetrating photons as they cross under to adjacent pixels. Presumably
    by the time that happens, the highlights are well blown out anyways.

    --
    Paul Furman
    www.edgehill.net
    www.baynatives.com

    all google groups messages filtered due to spam
     
    Paul Furman, Jul 21, 2008
    1. Advertisements

  2. John P Sheehy

    ejmartin Guest

     
    ejmartin, Jul 21, 2008
    1. Advertisements

  3. John P Sheehy

    ejmartin Guest

    This of course assumes that there are no noise sources that are
    constant as a function of pixel size. As you say below, reset noise
    might be one of those. Is CDS 100% efficient, or only up to a few
    electrons? And transistor noises at the pixel don't scale, according
    to the summary you linked. So we're agreed *if* it can be shown that
    there are no noise sources that don't scale.
    The calculations I did earlier in the thread show that the FZ50 does
    about a half stop better per area than the 1D3 at base ISO using the
    off-photosite limited DR figures that the 1D3 delivers. If one uses
    the full sensor DR of the 1D3, the FZ50 pixels will have to have read
    noise in the range of 1 electron, input referred, in order to match
    the full sensor DR of the 1D3 when it's not crippled by downstream
    electronics. Such a pixel is very close to a photon-counting device,
    which would be great if it could be achieved, but that means that all
    the potential noises from the pixel must be eliminated down to the
    single electron level. I would be pleasantly surprised if that were
    possible.


    Also, a followup point to the comment about capacitance dropping out
    of pixel DR in your model; it means the DR is independent of fill
    factor (only affecting QE I suppose), which also doesn't sound right
    to me.
     
    ejmartin, Jul 21, 2008
  4. John P Sheehy

    Bob Newman Guest

     
    Bob Newman, Jul 21, 2008
  5. John P Sheehy

    Bob Newman Guest

    Why should DR have anything to do with fill factor?
     
    Bob Newman, Jul 21, 2008
  6. John P Sheehy

    John Sheehy Guest

    It doesn't really seem to be a problem there, either. P&S cameras that
    do RAW (you can't tell a thing about sensors from cameras that do JPEG
    only) suggest that QE is not taking a toll from high pixel count, and
    pixel-level read noise doesn't seem to be on much of a rise, if any, so
    image-leve read noise is maintaining or improving.

    All the anecdotes about more noise from higher pixel density in P&S
    cameras come from viewing at 100%, or impressions made from NR designed
    for 100% pixel views.


    --
     
    John Sheehy, Jul 21, 2008
  7. John P Sheehy

    ejmartin Guest

     
    ejmartin, Jul 22, 2008
  8. John P Sheehy

    ejmartin Guest

    Though perhaps what you'll tell me is that the readout frequency goes
    up in inverse proportion to the pixel area if we want to keep the
    frame rate constant.
     
    ejmartin, Jul 22, 2008
  9. John P Sheehy

    Steve Guest

    Personally, I don't care much about just the sensor. I care about the
    entire image. So I can tell something from cameras that just do JPEG.
    If the camera trades off image softness for NR, resolution for S/N,
    etc., that will be obvious in the JPEGs also.
    I agree that a fairer comparison is to take a particular picture size,
    say 8"x10", and compare the 6MP P&S to the 12MP P&S at that size. That
    lets the higher MP camera take advantage of those pixels in the NR
    algorithms. I'm still not sure that a small sensor at 12MP can
    compete when you start pushing the ISO up to, say 400 or 800, with a
    6MP P&S in terms of image noise though, even if you do have those
    extra pixels.

    Steve
     
    Steve, Jul 22, 2008
  10. John P Sheehy

    ejmartin Guest

    A couple more references:

    www.imagesensors.org/Past%20Workshops/2007%20Workshop/2007%20Papers/056%20Leyris%20et%20al.pdf

    www-isl.stanford.edu/~abbas/group/papers_and_pub/1_f_noise.pdf
     
    ejmartin, Jul 22, 2008
  11. John P Sheehy

    Ray Fischer Guest

    You are the one who asked about number of pixels. Did you forget?
    But you made no mention of noise.
    You really are being an idiot since I never said otherwise. I said
    that people don't care about noise PER SENSOR AREA. Of course noise
    counts as detailed reviews on DPReview show. People DON'T care about
    noise per sensor area.
     
    Ray Fischer, Jul 22, 2008
  12. Not mentioning noise was intentional. I think that people actually care
    most about the final image results, not noise per pixel, or noise per
    sensor area. To understand /why/ an image is pleasing or not could
    involve either of those parameters, though.

    David
     
    David J Taylor, Jul 22, 2008
  13. John P Sheehy

    Bob Newman Guest

    Busy day today, won't get much chance to look at this (earning money
    gets in the way sometimes) but interesting article. Playing the
    'glances though and select quotes supporting my argument' game, we
    have:
    "In addition, increasing the V/e- conversion gain will further reduce
    noise proportionally. For example, the 2 e- noise floor demonstrated
    by the 5T CMOS imager described previously is generated by 50 µV/e-
    pixels. The conversion gain can be made as high as 200 µV/e- with 0.18
    µm CMOS design rules ", which seems to support my argument somewhat.
    I think the truth is that so far as read transistor noise is
    concerned, I am essentially on the right lines. However, there is a
    point at which 1/f noise and other sources which don't scale with V/e-
    will become significant and produce the 'noise floor' this article
    talks about. At that point, small pixels will cease to be better. What
    we don't know, is how far away from that we are. However, none of our
    current commercial imagers is close to sub e- read noise, so it looks
    like there's a bit to go yet. (Incidentally, JPS suggested he had
    picked up somewhere that sub e- read noise is possible, and was
    lampooned for it (not by you, I hasten to add) - this article
    corroborates that statement).
     
    Bob Newman, Jul 22, 2008
  14. John P Sheehy

    John Sheehy Guest

    (Ray Fischer) wrote in

    Well, they should, if they have in mind a camera with a certain sized
    sensor, which people often do. In that context, being concerned with
    noise per pixel would be foolish:

    "I've decided to buy a FF DSLR."

    "Oh; nice. Which one are you going to buy?"

    "The one that looks cleanest at 100% pixel view on my 96 PPI monitor, of
    course, because I am very slow upstairs."

    "Why not get the one that shows the best combination of resolution and
    image noise, printed at a standard size?

    "No one cares about quality per unit of area."

    --
     
    John Sheehy, Jul 23, 2008
  15. John P Sheehy

    Ray Fischer Guest

    Don't try to feed me bullshit in order to justify your weak argument.
     
    Ray Fischer, Jul 23, 2008
  16. John P Sheehy

    Bob Newman Guest

    That's a great reference. A bit more work on this, and we'll know a
    lot more about it.
    Meanwhile, from one of your references in the first post:
    "The white noise in terms of electrons at the CCD sense node is
    equivalent to the white noise in volts, defined above, divided by the
    product of amplifier sensitivity and output gain, as follows:

    Nwhite (rms electrons) = (4kTBRout)1/2 / Samp • Aamp
    In the above equation, S is the amplifier sensitivity and A represents
    the amplifier gain. The sensitivity value is a function of the
    fundamental electron charge and the CCD sense capacitance and is
    expressed in units of volts per electron."

    I think that just about confirms my input referred noise theory. The
    real issue now is the importance of the other noise sources at current
    sensor dimensions. Ultimately, when these become dominant, there will
    be a point of diminishing returns. We just don't know where it is
    yet.
     
    Bob Newman, Jul 23, 2008
  17. John P Sheehy

    ejmartin Guest

    Yes, after skimming these various articles it sounds as though you
    have the right model for that particular noise contribution, and the
    question is now a practical one about the magnitudes of other noise
    sources which provide a cutoff on the scaling behavior wrt pixel
    size. Some of those seem to be dependent on materials aspects like
    impurities. There are a few more articles about prototype devices and
    modelling noise sources at Eric Fossum's website

    http://imagesensors.org/Past Workshops/Past Workshops.htm

    It seems that 2 e- is about the limit of current prototype devices,
    and pixel sizes are edging toward 1µ but with more noise (see Micron's
    blurb in the 2007 workshop for the latter, and another article in the
    2007 workshop for the former, achieved by massive column-parallel
    readout). So perhaps current cameras are not close to fundamental
    limits.

    It does however seem to be the case that the best Canon can achieve
    with their current approach (and what they're willing to spend on fab
    costs per chip) is ~4 e- for read noise, a figure that has been rather
    constant since the 20D over a range of pixel sizes from 8.2µ (1D2) to
    5.7µ (40D) but we'll see what transpires with the impending small-
    pixel, small-sensor offerings that are supposedly imminent.
     
    ejmartin, Jul 23, 2008
  18. John P Sheehy

    ejmartin Guest

    BTW, something puzzles me -- if the cutoff is imposed by 1/f type
    noise sources, won't those be decreased by reading out the pixel
    faster? Which confuses me, I've always understood that read noise is
    reduced by slower readout (or was that a limitation of ADC's?), and is
    counter to the efforts to reduce read noise by using column parallel
    readouts that increase read times.
     
    ejmartin, Jul 23, 2008
  19. John P Sheehy

    Bob Newman Guest

    I think that there are two effects working against each other. High
    speed will limit 1/f noise, but capture chains tend to perform worse
    at high speed. The problem tends to be 'settling time'. Every stage
    has a limited output current and some capacitance. If you want to
    achieve a good sample, you need to sample it when that capacitance is
    fully charged. Push too much against the speed limits, or cut the
    drive current too much, and you tend to sample while it's still
    charging, and then you get jitter noise, due to irregularities in the
    sampling period. This is why, I suspect, the D3 A to D chain is so
    massively overspecified (one of those six AFE chips provides
    sufficient bandwidth according to the specs). The best answer would be
    to use a very high specced, high drive sample and hold (which actually
    determines the read time), which can capture the sample in a very
    small window, and thus minimise 1/f noise, followed by a nice
    leisurely ADC. I suspect that is what Sony is doing, and possibly
    Canon too. The two papers I've seen (on the D10 sensor and the 52MPix
    showcase) have on sensor sample and hold. I think the real advantage
    of the Sony column ADC architecture is not mainly noise, but cost. The
    very low spec ADC's can be integrated quite easily onto the sensor
    chip, and it saves a costly AFE off chip. Higher spec ADC's would
    require a costly BiCMOS process, which might not be available on CIS
    lines.
     
    Bob Newman, Jul 23, 2008
  20. John P Sheehy

    Bob Newman Guest

    I think that there are two effects working against each other. High
    speed will limit 1/f noise, but capture chains tend to perform worse
    at high speed. The problem tends to be 'settling time'. Every stage
    has a limited output current and some capacitance. If you want to
    achieve a good sample, you need to sample it when that capacitance is
    fully charged. Push too much against the speed limits, or cut the
    drive current too much, and you tend to sample while it's still
    charging, and then you get jitter noise, due to irregularities in the
    sampling period. This is why, I suspect, the D3 A to D chain is so
    massively overspecified (one of those six AFE chips provides
    sufficient bandwidth according to the specs). The best answer would be
    to use a very high specced, high drive sample and hold (which actually
    determines the read time), which can capture the sample in a very
    small window, and thus minimise 1/f noise, followed by a nice
    leisurely ADC. I suspect that is what Sony is doing, and possibly
    Canon too. The two papers I've seen (on the D10 sensor and the 52MPix
    showcase) have on sensor sample and hold. I think the real advantage
    of the Sony column ADC architecture is not mainly noise, but cost. The
    very low spec ADC's can be integrated quite easily onto the sensor
    chip, and it saves a costly AFE off chip. Higher spec ADC's would
    require a costly BiCMOS process, which might not be available on CIS
    lines.
     
    Bob Newman, Jul 23, 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.