From: Paul Rubin on
"acl" <achilleaslazarides(a)yahoo.co.uk> writes:
> > Is there experimental validation for this claim? My experience has
> > been not so encouraging but I'm probably not using the best possible
> > methods.
>
> If the only error is shot noise, then there is no difference ...
> What may be a problem is if there is a constant amount of noise per
> pixel which does not scale, then, the more you bin, the worse it becomes.

Precisely. Thus the question about experimental validation. I think
we've pretty much established that with large pixels, shot noise
dominates. I don't think this has been established for small pixels
From: David J. Littleboy on

"Lionel" <usenet(a)imagenoir.com> wrote:
> On Fri, 16 Mar 2007 20:40:49 GMT, John Sheehy <JPS(a)no.komm> wrote:
>>"David J. Littleboy" <davidjl(a)gol.com> wrote:
>>
>>> No. I think what John is saying is orthogonal to sensor size
>>> arguments. He's arguing that for a given sensor size, one wants as
>>> many pixels as one can get. Roger is arguing that for a given number
>>> of pixels, one wants the largest sensor you can get.
>>>
>>> I suspect that they're both right.
>>
>>Yes, if that is what Roger is arguing. There are legions of people,
>>however, in other web forums that quote Roger as proof that bigger pixels
>>are always better, in all situations, and that the "megapixel race" is a
>>race to nowhere. He could be clearer if he means what you say, but I get
>>the impression that he does believe that subdividing a given pixel real-
>>estate into smaller pixels lowers the bottom line in image quality.
>
> It does, because of the fill-factor problem. I've gone into this in
> more detail in my reply to another of your posts in this thread, but
> the bottom line is that the more pixels you put in a given area, the
> more of that area is being used for things other than collecting
> photons. In fact, as you increase the number of pixels in an area, you
> eventually reach the point where the whole area is support
> electronics, & there's no longer any room left for the actual
> photodiodes.

Not so much to argue with you, but to get clarification from people who know
more of the details than I do.

(One thing that you don't mention, but makes your point more valid, is that
I suspect that smaller microlenses are harder to fabricate and probably not
as efficient as larger ones.)

How much of an issue is fill factor _in the presence of microlenses_? I'd
think that microlenses would mean that fill factor in the silicon itself is
much less of an issue. (This is the main issue I'd like to see addressed.)

Also, my understanding (possibly wrong) is that the CCD/CMOS sensors are
fabricated in technology that's multiple generations behind the
DRAM/microprocessor/ASIC curve, so there's actually quite a bit of room for
improvement in the circuits.

My (jaundiced as usual) reading between the line of the
FillFactory/Kodak/Dalsa stuff was that these guys don't (or didn't) have the
technology to fabricate microlenses and thus the fill factor really is/was a
big deal for these guys.

David J. Littleboy
Tokyo, Japan


From: acl on
On Mar 17, 5:55 am, Paul Rubin <http://phr...(a)NOSPAM.invalid> wrote:

> Precisely. Thus the question about experimental validation. I think
> we've pretty much established that with large pixels, shot noise
> dominates. I don't think this has been established for small pixels

Well for high signal levels I certainly hope it does. But what
dominates per pixel isn't the point, the point is that if we read the
pixels, each having read noise r, and bin them nxn, the read noise per
binned pixel will be effectively n*r. So we'd need lower read noise.

For example, if we take the data Roger has for the S70 and the D200
here
http://www.clarkvision.com/imagedetail/digital.sensor.performance.summary/index.html
we see that if we binned the image from the S70 3x3 we'd get
effectively around 9 electrons read noise (as opposed to around 7 for
the D200; I assume these are at unity gain which is different, but
never mind, suppose the read noise is constant). But shot noise would
be potentially less. So a larger version of the S70 sensor, with the
same-size pixels (I think this would be around 20mpx) could perform
close to a D200 by binning, or with twice the pixel and worse low-
light capabilities. This is because it has less read noise (per
pixel).

If this does not work when someone tries it (I don't have an S70, and
the other compact there seems to have much higher read noise), either
Roger's data isn't correct (which I doubt), or the noise is
correlated.

From: acl on
On Mar 17, 6:48 am, Lionel <use...(a)imagenoir.com> wrote:
> On 16 Mar 2007 20:33:30 -0700, "acl" <achilleaslazari...(a)yahoo.co.uk>
> wrote:
>
> >On Mar 17, 5:55 am, Paul Rubin <http://phr...(a)NOSPAM.invalid> wrote:
>
> >> Precisely. Thus the question about experimental validation. I think
> >> we've pretty much established that with large pixels, shot noise
> >> dominates. I don't think this has been established for small pixels
>
> >Well for high signal levels I certainly hope it does. But what
> >dominates per pixel isn't the point, the point is that if we read the
> >pixels, each having read noise r, and bin them nxn, the read noise per
> >binned pixel will be effectively n*r. So we'd need lower read noise.
>
> Your read noise is going to be multplied by the number of photodiodes
> you're binning. Worse, because the signal level is proportional to the
> size of the photodiode, your read noise for each pixel increases (ie;
> the signal to read noise ratio decreases) with decreases in photodiode
> size.
>

No it's multiplied by the sqrt of the number of pixels you're binning.
This is exactly what I say in the part you quoted. I also do not see
why the absolute value of the read noise will decrease with decreasing
size. True, the s/n decreases because the signal decreases, but blah
blah. Please, read the post to which you replied a bit more carefully,
I think that would be the answer to what you say.

From: acl on
On Mar 17, 6:13 am, Lionel <use...(a)imagenoir.com> wrote:
> On Sat, 17 Mar 2007 02:21:02 GMT, John Sheehy <J...(a)no.komm> wrote:

> >Even if you lose some photons due to miniaturization, remember, the shot
> >noise is only proportional to the square root of the signal.
>
> IIRC, it's proportional to the temperature of the sensor, & the bias
> across the photodiode, which is not the same thing. When we start

Shot noise in this thread refers to fluctuations in the number of
detected photons due to the random intervals between their arrival
(hence poissonian); it has absolutely nothing to do with temperature
etc. There are other noise sources that do.

> discussing things in this sort of detail, hand-waving & analogies stop
> being useful, & we need to refer to the underlying physics & the
> characteristics of real photodiodes.

These aren't hand-waving analogies, it's the real thing. The only
quantum aspect of the photons relevant to shot noise here is that
they're discrete particles.

> There's a very good reference
> available from one manufacturer of photodiodes at:
> <http://sales.hamamatsu.com/assets/applications/SSD/photodiode_technic...>
> You'll find the noise characteristics in section 2-4. Do bear in mind

This is about read noise and similar noise sources (the ones that will
be a problem in this scenario, indeed).

> I think that where you're going wrong is that you're forgetting that
> with the really tiny signals we're discussing, popular, macro-level
> analogies for light & electricity stop making sense, & we have to look
> at the quantum characteristics of photons & electrons. See, you can't
> measure half a photon, or a third of an electron. Like a brick, you
> either see it or you don't. if you replace the raindrops in your
> analogy with bricks, & imagine that it's raining bricks (in proportion
> to the temperature!) at the same time, then the problem with trying to
> count the bricks (photons/electrons) in smaller buckets (photodiodes)
> is a /lot/ clearer. ;^)

That is exactly why what he says works. The problem is a) a fixed
error per bucket, so they add up (or their squares do), b) nxn smaller
buckets don't cover the same area as a larger bucket of area n^2 (fill
factor<1). This is sort of what you're saying, but it does not
contradict what he is saying.