From: Paul Rubin on
John Sheehy <JPS(a)no.komm> writes:
> Your assistant is me, and I decide to replace the 16 containers with 64
> square containers, 4 of which fit in the same space as 1 of yours. I come
> back to you with a list of results from 64 smaller containers, instead of
> the 16 you asked for. The list is longer, and the total count is the
> same as it would be if I used your original containers, but have I
> created any *NOISE*?

Well, how to you measure how many raindrops fell in each container?
Let's say exactly 100 raindrops fell in each of the original 16
containers. Your counting method is accurate to within +/- 5
raindrops, so you get counts from 95 to 105 inclusive, not too bad for
this type of measurement; each count is within a 5% error band.

Now you use the 64 smaller containers, and your counting method is
still accurate to +/- 5 raindrops. But now you only have 25 raindrops
per container on average, so now your +/- 5 raindrop error is a 20%
band. And the total count for each 4 raindrop cluster can be anywhere
from 80 to 120 instead of 95 to 105.

In this case I would say you have added more noise by using smaller
containers, since you haven't decreased the counting error per container.
You've taken a +/- 5% accurate measurement and replaced it with a +/- 20%
measurement.
From: Paul Rubin on
"David J. Littleboy" <davidjl(a)gol.com> writes:
> The point/claim is that pixel binning (or noise reduction plus downsampling)
> will result in the same image as the larger pixels would have in lower
> light. One is collecting the same number of photons, so this should/might
> work.

Is there experimental validation for this claim? My experience has
been not so encouraging but I'm probably not using the best possible
methods.
From: John Sheehy on
Lionel <usenet(a)imagenoir.com> wrote in
news:7gamv2h9ik0k7vrlqqbvgilb835c1f13p6(a)4ax.com:

> If fill-factor wasn't an issue, your reasoning would be correct.
> However, the more photodiodes you put in a given area, the more of
> that area is used for the support structures (eg; the array selection
> & blanking transistors) that are needed *for every photodiode* to get
> the data out of them & over to the amplifiers. (Using your analogy,
> think of them as the hoses, pumps & taps that go to each of your
> buckets.) There is a minimum practical size those structures can
> be[0], so the more pixels you have in a given area, the more of it is
> being used for those support electronics, rather than for detecting
> photons.

> [0] With the present state of the art in semiconductor fabrication.
> This will likely improve over time, but it's a slow process.

Yeah, this is what I believed, until I realized that the tiny-sensor ZLRs
were capturing as many photons per square mm of sensor area as big-pixel
DSLRs. Canon's on-site transistors are a two-edged sword; they help
greatly with relatively low read noise at high ISOs, but they get in the
way of minaturization. The 1.97u pixels from a Panasonic FZ50, filling a
full 36x24mm frame, would have 223MP. Imagine a very sharp lens, like the
500mm f/4L on that.

Even if you lose some photons due to miniaturization, remember, the shot
noise is only proportional to the square root of the signal. You gain a
small amount of shot noise, and a whole lot of resolution.

--

<>>< ><<> ><<> <>>< ><<> <>>< <>>< ><<>
John P Sheehy <JPS(a)no.komm>
><<> <>>< <>>< ><<> <>>< ><<> ><<> <>><
From: acl on
On Mar 17, 5:12 am, Paul Rubin <http://phr...(a)NOSPAM.invalid> wrote:
> "David J. Littleboy" <davi...(a)gol.com> writes:
>
> > The point/claim is that pixel binning (or noise reduction plus downsampling)
> > will result in the same image as the larger pixels would have in lower
> > light. One is collecting the same number of photons, so this should/might
> > work.
>
> 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 bet
binning 4 pixels and using a pixel twice as long. eg create a random
image (add noise to an empty image) in ps, measure the std and then
bin and repeat. Or in real images, find a uniform area and do the
same. It works (if it didn't, it would mean that the noise isn't
uncorrelated, and then you could model it and remove it). 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. Also the
finite fraction of pixel area taken up by non-photosensitive stuff.

From: David J. Littleboy on

"Paul Rubin" <http://phr.cx(a)NOSPAM.invalid> wrote in message
news:7xmz2cbots.fsf(a)ruckus.brouhaha.com...
> "David J. Littleboy" <davidjl(a)gol.com> writes:
>> The point/claim is that pixel binning (or noise reduction plus
>> downsampling)
>> will result in the same image as the larger pixels would have in lower
>> light. One is collecting the same number of photons, so this should/might
>> work.
>
> Is there experimental validation for this claim? My experience has
> been not so encouraging but I'm probably not using the best possible
> methods.

Note that I added the word "claim" in there<g>.

I just played with this ISO 25,600 5D image (taken at ISO 3200, underexposed
by three stops, and pushed in postprocessing to ISO 25,600). What I've done
is taken an ISO 25,600 image, blasted it with full-tilt Lightroom noise
reduction, and downsampled by a factor of 9. The idea is that a 1.5MP
full-frame image at ISO 25,600 ought to have similar pixel level noise to an
ISO 3200 12.5MP full-frame image.

Pushed with no other processing, 100% crop of 5D ISO 25,600 image.
http://www.pbase.com/davidjl/image/75374090/original

100% luma and chroma NR in Lightroom, 1.5-pixel Gaussian blur in Photoshop,
downsampled to 1.5MP, and sharpened. Full 1.5MP image.
http://www.pbase.com/davidjl/image/75751501/original

It's hard to compare this to the other images, since the scaling is so
different.

For comparison, here are a sequence of 5D images processed for printing as
best I could at ISO 3200, 6400, 12,800, and 25,600. I think the 6,400 image
would make a superb 8x10, and a surprisingly good 12x18. But the speckles
are out of control in the 12,800 image, and the 25,600 image was completely
unsharpenable.

http://www.pbase.com/davidjl/image/75359389/original

David J. Littleboy
Tokyo, Japan