Digital camera sensor performance

Discussion in 'Digital SLR' started by Roger N. Clark (change username to rnclark), Nov 19, 2006.

  1. I have a new page up describing digital camera sensor performance
    for a fair range in digital cameras and commercial sensors.
    Enough data has now been measured to show trends in full-well
    capacity, dynamic range, signal-to-noise ratios,
    unity gain ISO (high ISO performance), and how read noise
    does not follow the same trends.

    Included are some newly derived parameters showing low light performance.
    I've plotted results as functions of pixel size and there are clear trends.
    Some trends can be used to predict basic performance of new cameras.

    See:
    http://www.clarkvision.com/imagedetail/digital.sensor.performance.summary

    Roger
     
    Roger N. Clark (change username to rnclark), Nov 19, 2006
    #1
    1. Advertisements

  2. Thanks Roger ... that is very useful and well-done!
     
    Charles Schuler, Nov 19, 2006
    #2
    1. Advertisements

  3. Looks like I made the right choice when I ordered my 5D <g>
     
    Scott in Florida, Nov 19, 2006
    #3
  4. Roger N. Clark (change username to rnclark)

    Paul Furman Guest

    What are the KAI-11002... (kodak sensors) used for, old digital backs?
    In figure 7 they have awful low light sensitivity in spite of large pixels.

    Pixel pitch is the spacing, not the actual size, right? Wouldn't Full
    Well Capacity be the same as the actual pixel size? So in figure 1, the
    D50 has a small well for it's pitch, same for SNR but the unity gain in
    fig 6 is good for it's pitch, almost as good as the 5D.

    Are the Canon S60/70 much better than typical modern P&S or sort of average?

    Is it fair to say that pixel count effects SNR? As the grain gets
    smaller, it's certainly less noticeable but I don't know maybe that
    doesn't really count.
     
    Paul Furman, Nov 20, 2006
    #4
  5. Paul,
    I am not sure what the Kodak sensors are used in. They have good
    sensitivity for signals above a few tens of photons. Their
    low value in Figure 7 is due to their high read noise, double
    the Nikon CCDs and about 4 time the Canon CMOS sensors.
    Correct. Pixels have an active area that is somewhat
    smaller than the pixel pitch squared. The active area
    divided by the pitch squared is the fill factor. The fill factor,
    or effective fill factor with the micro lenses over each pixel
    is usually in the 80-90% range. So pixel pitch is a pretty
    good measure, and usually the only parameter us consumers
    can get for pixel size
    Yes, it must have a high Quantum Efficiency.
    The S60 is a couple of years old and probably pretty average.
    The S70 is probably also pretty average except that the
    read noise is about half that of other CCDs.
    Yes, that is a good point, and I think it does count, at least
    to some degree. After all, film has very low SNR, but
    with many grain clumps, like pixels, the finer they are,
    the less noticeable they become. But as you try and make
    bigger enlargements, the warts show.

    Roger
     
    Roger N. Clark (change username to rnclark), Nov 20, 2006
    #5
  6. Roger N. Clark (change username to rnclark)

    acl Guest


    Roger, thanks for all the hard work you've put into this. Do you have
    any idea why the 5D and 20D have such high read noise in low ISOs? I
    imagine it will not be a pleasant experience to significantly lift the
    shadows in an ISO 100 image from them!

    Still, it dramatically decreases with increasing ISO. I would be
    extremely interested to find out how Canon do this (not that they'll
    tell us any time soon). Very impressive.
     
    acl, Nov 22, 2006
    #6
  7. While cameras like the 5D, 20D, and 1D Mark II have apparently higher
    low ISO read noise in electrons, the image noise is low. The read noise
    is high because of A/D quantization and a small amount of noise
    in the amplifier. For example, the 20D has a read noise of only 3.6
    at ISO 1600 but 25 electrons at ISO 100. But in terms of data numbers
    (DN) in the image, that 25.3 electrons becomes a noise of 2.0 DN.
    Thus, there is amazing detail in the shadows of an image. One may also
    have some fixed pattern noise, but that can be calibrated out.
    (Take the read noise, Table 4, and divide by the gain in Table 3
    to get the data numbers in an image.)

    Roger
     
    Roger N. Clark (change username to rnclark), Nov 22, 2006
    #7
  8. Roger N. Clark (change username to rnclark)

    acl Guest

    Yes, that's what I did. The D200, for instance, would appear from these
    numbers to have better shadow detail at ISO 100 (its read noise
    corresponds to 1.25 DN as opposed to around 2DN for both the 5D and the
    20D at the same ISO; things change a lot at higher ISOs!).

    At any rate, read noise of this magnitude (DN 1 or 2) is not going to
    affect the image too much.

    My question, though, is if you have any idea how they manage to decrease
    read noise so much when ISO is increased (as unity gain is approached,
    judging from the compacts that you have also tested). I am not an expert
    in imaging systems, but this seems almost like magic to me.
     
    acl, Nov 22, 2006
    #8
  9. But fixed pattern noise can easily be larger than a DN or so.
    So to say which is better needs images from both cameras
    to be evaluated. Then note that the larger pixels of
    the 5D versus D200 would result in more photons per pixel for the
    same f/ratio lens so the 5D image would have a signal/noise
    advantage boosted by about 2x.
    I have heard something about double correlated reads, but have not
    been able to confirm that is what Canon actually does. Another
    speculation is something about the reset cycle reducing fixed
    noise and minimizing electrons in the capacitor on each pixel
    to reduce the read noise. But without a technical paper from
    Canon, one can't be sure. I have not found a Canon technical
    paper, and perhaps it is because Canon recognizes they have
    an advantage and want to keep it to themselves.

    Roger
     
    Roger N. Clark (change username to rnclark), Nov 28, 2006
    #9
  10. Roger N. Clark (change username to rnclark)

    Lionel Guest

    Lionel, Nov 28, 2006
    #10
  11. Roger N. Clark (change username to rnclark)

    Lionel Guest

    One trick might be to improve readout accuracy by pre-biasing the
    photosites with a charge proportional to the ISO setting, then
    subtracting the bias value in software after conversion.
    Now that I think about it, that might also reduce the effect of
    thermal noise.
     
    Lionel, Nov 28, 2006
    #11
  12. Roger N. Clark (change username to rnclark)

    acl Guest

    I see. That would decrease the max number of electrons that can be held
    before dielectric breakdown, no? So decrease max well capacity. It's
    like throwing away your shadow range, isn't it? Except you just push it
    up, throwing away highlight range instead. Do I have this right?

    But can this be done for CCDs? (I have no clue, I am not an engineer;
    it's a question).
     
    acl, Nov 28, 2006
    #12
  13. Roger N. Clark (change username to rnclark)

    acl Guest

    It wouldn't reduce thermal noise, though (I think).
     
    acl, Nov 28, 2006
    #13
  14. That would add noise. Any signal has an uncertainty equal
    to the square root of the number of electrons.

    Roger
     
    Roger N. Clark (change username to rnclark), Nov 29, 2006
    #14
  15. I think his idea is that low electron counts are not counted
    accurately, so let's avoid them. If that's the case, it would work by
    avoiding low electron counts being read. I have no clue if it is so.

    I don't know if what he says works, but I am pretty sure he did not
    mean it as a way to avoid shot noise (indeed, it will add its own
    noise, as you say).
     
    achilleaslazarides, Nov 29, 2006
    #15
  16. Roger N. Clark (change username to rnclark)

    Lionel Guest

    Of course, but if it adds *less* noise than the extra noise you get
    from the extra amplification required to bump up the effective ISO of
    the sensor, then it's a net gain. I don't have the data required for a
    rigourous analysis, (or any realistic way of obtaining it, short of
    being hired by Canon for an R&D role), but my gut instinct is that it
    would help a lot - particularly in reducing the influence of thermal
    noise (eg; coloured hot pixels on Bayer sensors) on low-light images.
     
    Lionel, Nov 30, 2006
    #16
  17. But increasing the ISO does not have the same effect as adding a
    constant to the measured value. Also, even if you could add this value
    and subtract it without any noise, the only advantage you will get will
    be to avoid any problems related to quantising small charges. This is
    what I thought you meant. You will not avoid problems such as a small
    actual signal being drowned out by noise (since the signal is still
    small).

    If you meant something else, I can't think what it can be (and still
    have a chance of being right). Can you explain?
     
    achilleaslazarides, Nov 30, 2006
    #17
  18. Roger N. Clark (change username to rnclark)

    Lionel Guest

    I have no hard evidence that it would either, I'm merely making an
    educated guess - albeit one based on my technical experience with
    data-aquisition design (which I'll cheerfully admit is dated, but
    still broadly valid), & my subjective experience with visible noise in
    images captured with Canon CMOS image sensors.
    Correct. I'm talking about a tradeoff off (virtual) well-depth for amp
    noise at very low signal levels. I believe the concept is reasonably
    sound in theory, but its usefulness in practice would depend a lot on
    the noise & thermal characteristics of the particular sensor & read
    amps. That said, don't forget that DLSR image sensors (& their amps!)
    have to work at temperatures much higher than optimum, so thermal
    noise is very likely one of the mjor design issues, & anything that
    helps with that would be valuable.
     
    Lionel, Nov 30, 2006
    #18
  19. Roger N. Clark (change username to rnclark)

    acl Guest

    I just saw this post! OK, this makes sense then. Ignore my other post.
     
    acl, Nov 30, 2006
    #19
    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.