From: Roger N. Clark (change username to rnclark) on
ejmartin wrote:
> On Jul 17, 10:56 pm, "Roger N. Clark (change username to rnclark)"
> <usern...(a)qwest.net> wrote:
>> John,
>> Your post is full of inaccuracies. I'll hit a few of the
>> highlights.
>
>> The problem is on the bottom end. Let's say the pixel read noise is about
>> the same regardless of pixel size (which is true in real world sensors).
>> Let's say the pixel density of the small pixel sensor is 100 times the
>> large pixel sensor. The total read noise from the 100 small pixels
>> results in a signal over those 100 pixels to equal the
>> large pixels is 10 times (square root 100) the single pixel.
>> So in the end you lose. The noise floor is 10x higher than the single
>> pixel, thus you lose both sensitivity and dynamic range.
>> Before you think you can correct me on this, read below
>> where I show your math mistake.
>
> I fully agree about read noise; but I'm not sure that John is making
> the elementary math errors you infer. In part this is because John
> hasn't fully explained the context of his various measurements here,
> though he has in various threads over at DPReview.

John's math error, e.g.:

John Sheehy wrote on 7/17/2008:
> Shot noise per unit of area is lower in the FZ50, for a given exposure.
> It has one of the highest QEs in the industry. Per pixel, the 400D
> collects up to about 43K photons 4 stops above metered middle grey, the
> FZ50, about 4800 photons 2.5 stops above metered middle grey. 4800*
> 83.521 = 40,090, so the FZ50 sensor collects a maximum number of photons
> per unit of area slightly less than the 400D.

So there he summed the signal from the small pixels, then in the next
paragraph:

> Read noise per pixel is about 2.8 12-bit ADU in the FZ50, and about 1.65
> 12-bit ADU in the 400D, both at ISO 100. Scaled by pixel pitch, the read
> noise of the FZ50 is about 2.8/2.89 = 0.97 ADU, in 400D pixel terms.

He averages the read noise. So he has mixed two equations, summing
the signal and averaging the read noise. That is wrong.
Plus he adds confusion of working in mixed electrons and ADU,
although one is proportional to the other.

Then the next day he corrected his math error in new examples, but
now uses future technology to better existing consumer technology:

John Sheehy wrote 7/18/2008:
> I've heard from more than one source
> that read noises can be had as low as 1 electron with 1-micron pixels, in
> the lab. So, let's say that each 1-micron pixel collected up to 1000
> electron charges; that's a ratio of 1000:1 for max signal to read noise.
> 4 of these pixels would be 4000:2 or 2000:1, 9 would be 9000:3 or 3000:1
> (about the same as the 1-series Canons and the D3 at the virtual pixel
> level, and you're still only binning to 3 microns). At 8 microns, you
> have 64000:8, or 8000:1, a virtual pixel-level DR far beyond *any* DSLR
> with ~8 micron pixel pitches.

But then there is no current evidence that there is a dependence of read
noise with pixel size; see Figure 3 at:
http://www.clarkvision.com/imagedetail/digital.sensor.performance.summary

>
> I think John's error lies in using low ISO read noise figures to infer
> the properties of the photosites, which incorrectly inflates the read
> noise of DSLR's relative to digicams. The former have low ISO DR
> which is limited by the DR of the downstream circuitry, while the
> latter have so little photosite DR that the downstream processing can
> easily accomodate it.

I agree.

> It does seem quite high, but I think I'm with John on this one. I did
> a crude analysis of a publicly available raw file:
>
> http://www.rawsamples.ch/raws/panasonic/fz50/RAW_PANASONIC_FZ50_SRGB.RAW
>
> since I don't have access to an FZ50. Using the difference of the two
> green subarrays as an ersatz difference file, the standard deviation
> relative to the mean of uniform bright areas yielded a gain of 1.2
> electrons/12-bit ADU. Note that this is a lower bound, since by
> subtracting values from two different pixels there are other noise
> sources that contribute besides just photon shot noise. Since raw
> saturation is at 4095, this is consistent with John's figure of 4800
> electrons at raw saturation.

It is not valid. There are many examples of people extrapolating
in error single measurements like this. The 2 red flags are:
1) QE so high only custom thinned back side illuminated sensors have
that level of performance, and
2) Electron densities are higher than any other reported number.

Those to together imply significant errors in the full well and gain.
High QE, thinned back-side illuminated sensors tend to have lower full
wells, also a red flag of the two above conditions.

> There has been a heated discussion over at DPReview forums, where the
> proprietor of the review site has introduced a 'figure of merit' for
> cameras, the pixel density -- pixel count divided by sensor area. He
> thinks it is an appropriate measure of image quality.

I also do not agree that "pixel density" is a "figure of merit."
While it is a measure, it is only one indicator in a larger equation,
which is what I think John is trying to point out. unfortunately,
John also is not seeing the bigger picture (pun intended).
Probably a better "figure of merit" would be:

Clarkvision figure of merit (CFM) = megapixels * pixel pitch. ;-)

Pixel pitch being proportional to the square root of the pixel density,
and pixel pitch is also related to signal-to-noise ratio, which is
a main property of my AIQ function. But the above equation does
ignore QE, filter transmission and fill factor variations which is better
represented in the S/N in my AIQ equation. However, we do not have
raw data and S/N sensor info for many cameras, so CFM may be a good
overall indicator. I've added that to my sensor performance summary web
page and I'll expand it when I get some time.
However, this CFM ;-) equation does ignore
the small pixel effects of vanishing DR as pixel size decreases, and
lower image detail due to diffraction effects, which is in my AIQ model.
It also ignores lens quality, but so does simple pixel pitch.

Some examples:
Camera Megapixels Pixel Pitch CFM
(MP) (microns) (MP*microns)
Canon 1Ds Mark III 21.1 6.4 135
Nikon D3 12.1 8.46 102
Canon 1D Mark II 8.2 8.2 67
Canon 40D 10.1 5.7 57
Canon 30D 8.2 6.4 52
Olympus E3 10.0 4.7 47
Panasonic FZ50 10.0 1.97 19.7
Canon S70 7.1 2.3 16.3
Panasonic FZ18 8.1 1.76 14.3


Roger
From: ejmartin on
On Jul 19, 3:50 pm, "Roger N. Clark (change username to rnclark)"
<usern...(a)qwest.net> wrote:
> I've added that to my sensor performance summary web
> page and I'll expand it when I get some time.
> However, this CFM ;-) equation does ignore
> the small pixel effects of vanishing DR as pixel size decreases, and
> lower image detail due to diffraction effects, which is in my AIQ model.
> It also ignores lens quality, but so does simple pixel pitch.
>
> Some examples:
> Camera              Megapixels  Pixel Pitch   CFM
>                         (MP)      (microns)  (MP*microns)
> Canon 1Ds Mark III    21.1         6.4       135
> Nikon D3              12.1         8.46      102
> Canon 1D Mark II       8.2         8.2        67
> Canon 40D             10.1         5.7        57
> Canon 30D              8.2         6.4        52
> Olympus E3            10.0         4.7        47
> Panasonic FZ50        10.0         1.97       19.7
> Canon S70              7.1         2.3        16.3
> Panasonic FZ18         8.1         1.76       14.3
>

A bit OT, but as long as you're updating your webpages, one thing I'd
like to see is the main data table (Table 2) re-sorted by Manufacturer
(All Canons in one block, all Nikons together, etc) and then within
that, ordered by camera class and tech generation (eg Canon 10D, 20D,
30D, 40D together; 1D, 1D2, 1D3 together, etc). Currently it's a pain
having to search through the whole list to find a particular camera,
since even contemporaneous cameras (Nikon D3 and D300) are in
completely different locations. OK, end of rant ;^)

From: ejmartin on
On Jul 19, 3:50 pm, "Roger N. Clark (change username to rnclark)"
<usern...(a)qwest.net> wrote: 
>
> I also do not agree that "pixel density" is a "figure of merit."
> While it is a measure, it is only one indicator in a larger equation,
> which is what I think John is trying to point out.  unfortunately,
> John also is not seeing the bigger picture (pun intended).
> Probably a better "figure of merit" would be:
>
>      Clarkvision figure of merit (CFM) = megapixels * pixel pitch..  ;-)
>
> Pixel pitch being proportional to the square root of the pixel density,
> and pixel pitch is also related to signal-to-noise ratio, which is
> a main property of my AIQ function.  But the above equation does
> ignore QE, filter transmission and fill factor variations which is better
> represented in the S/N in my AIQ equation.  However, we do not have
> raw data and S/N sensor info for many cameras, so CFM may be a good
> overall indicator.  

All right, back on topic. In your AIQ, is the SNR component measured
at the pixel level (Nyquist frequency) or at some fixed spatial
frequency relative to frame size? I would think the latter is
appropriate; the SNR at the pixel level is strongly correlated to the
pixel pitch, but the image SNR averages pixel level values out to a
fixed spatial frequency, a fixed scale relative to frame size.
However, your insertion of pixel pitch in place of SNR in generating
this new measure leads me to believe you are using pixel level noise
figures and pixel level SNR's. Noise comparisons between cameras of
different pixel pitch should use the pixel level noises rescaled by
the pixel pitch, so that they are compared at the same spatial
frequency. That more closely reflects the viewer's experience looking
at two such images side-by-side, where the eye takes in image data
referred to the frame size, not to the pixel size. Using pixel level
values amounts to continually changing the viewer's distance from the
image so that a pixel is always the same viewed solid angle relative
to the eye, and unfairly disadvantages small pixels by zooming in to
the image and magnifying the noise that is there on small scales.

This is I believe part of what John is trying to get at, and in this
respect I agree with him.
From: Bob Newman on
On Jul 19, 9:50 pm, "Roger N. Clark (change username to rnclark)"
<usern...(a)qwest.net> wrote:
> ejmartin wrote:
> > On Jul 17, 10:56 pm, "Roger N. Clark (change username to rnclark)"
> > <usern...(a)qwest.net> wrote:
> >> John,
> >> Your post is full of inaccuracies. I'll hit a few of the
> >> highlights.
>
> >> The problem is on the bottom end. Let's say the pixel read noise is about
> >> the same regardless of pixel size (which is true in real world sensors).
> >> Let's say the pixel density of the small pixel sensor is 100 times the
> >> large pixel sensor. The total read noise from the 100 small pixels
> >> results in a signal over those 100 pixels to equal the
> >> large pixels is 10 times (square root 100) the single pixel.
> >> So in the end you lose. The noise floor is 10x higher than the single
> >> pixel, thus you lose both sensitivity and dynamic range.
> >> Before you think you can correct me on this, read below
> >> where I show your math mistake.
>
> > I fully agree about read noise; but I'm not sure that John is making
> > the elementary math errors you infer. In part this is because John
> > hasn't fully explained the context of his various measurements here,
> > though he has in various threads over at DPReview.
>
> John's math error, e.g.:
>
> John Sheehy wrote on 7/17/2008:
> > Shot noise per unit of area is lower in the FZ50, for a given exposure.
> > It has one of the highest QEs in the industry. Per pixel, the 400D
> > collects up to about 43K photons 4 stops above metered middle grey, the
> > FZ50, about 4800 photons 2.5 stops above metered middle grey. 4800*
> > 83.521 = 40,090, so the FZ50 sensor collects a maximum number of photons
> > per unit of area slightly less than the 400D.
>
> So there he summed the signal from the small pixels, then in the next
> paragraph:
>
> > Read noise per pixel is about 2.8 12-bit ADU in the FZ50, and about 1.65
> > 12-bit ADU in the 400D, both at ISO 100. Scaled by pixel pitch, the read
> > noise of the FZ50 is about 2.8/2.89 = 0.97 ADU, in 400D pixel terms.
>
> He averages the read noise. So he has mixed two equations, summing
> the signal and averaging the read noise. That is wrong.
> Plus he adds confusion of working in mixed electrons and ADU,
> although one is proportional to the other.
Yes, this is confusing, but working in one unit only also causes
confusion. Particularly concerning 'read noise' in electrons, when
it's a voltage referred noise. This means that the read noise, which
is actually a proportion of output voltage of the sensor amplifier
chain, should scale down as pixel area when expressed as electrons,
since that output voltage represents a number of electrons scaled down
by the pixel area. Now when we add those noise contributions back
together to form the original pixel area, they add in quadrature to
give a lower noise (electron referred) than the single pixel.
Actually, by splitting the read into multiple channels, what has
occurred is that the noise in the channels is averaged and therefore
reduced, since it's uncorrelated between the channels.
The key question here is whether the read noise really is voltage or
charge referred. You claim that it's constant over pixel size, but
your own figures don't really support that - it's all over the place.
Neither is there the inverse relationship I suggest apparent. In
practice, I think that there is sufficient difference in the design of
pixels for individual sensors that you can't simply use these
measurements to support scale dependent arguments, either way.
> Then the next day he corrected his math error in new examples, but
> now uses future technology to better existing consumer technology:

> But then there is no current evidence that there is a dependence of read
> noise with pixel size; see Figure 3 at:http://www.clarkvision.com/imagedetail/digital.sensor.performance.sum...
Yes, but as I pointed out above, there's no evidence the other way,
either.

> I also do not agree that "pixel density" is a "figure of merit."
> While it is a measure, it is only one indicator in a larger equation,
> which is what I think John is trying to point out. unfortunately,
> John also is not seeing the bigger picture (pun intended).
> Probably a better "figure of merit" would be:
>
> Clarkvision figure of merit (CFM) = megapixels * pixel pitch. ;-)
>
An interesting figure of merit. For a given size sensor, the more
pixels you jam into it, the higher the CFM (since megapixels increases
as the inverse square of the pixel pitch) - perhaps about right.


From: John Sheehy on
Steve <steve(a)example.com> wrote in
news:vvj4845bi7rih5tkccjct293mtsg3gnmj5(a)4ax.com:

>
> On Sat, 19 Jul 2008 15:02:55 GMT, Steve <steve(a)example.com> wrote:
>
> [...]
>>However, using John's test, you can show that 1000 megapixels for an
>>APS-C sized sensor has better S/N and DR than the 10 megapixel sensor.

I have no idea what a GP would do to pixels with current technology. I
imagine that some changes would have to occur to have pixels that small.
A whole different technology. It's all up to what happens to QE and read
noise.

>>Because he would only use the portion of the DR that the 1000 MP
>>sensor can manage when comparing it to the 10MP sensor, which would
>>have tons of wasted headroom in his test.

Huh? Where is tons of headroom coming from? The difference in headroom
between two sets of pixels depends on the QE, and/or how the camera
meters. The QE will determine the noise quality of the higher ranges.

>> And he would cut a tiny
>>area of the focal plane image out of the 10MP sensor and blow it up to
>>10,000% and compare that with a 100% crop of the 1000MP camera.
>
> And on top of that, he would use the ISO setting that the 1000MP could
> work at for the entire test for both sensors and not vary the range if
> ISO to include values that normal photographers would want to use.
> Limiting his test to ISO 100 for both cameras and then pushing it to
> 13,500 is also silly. Lets see his results for ISO 400, ISO 800 and
> ISO 1600.

> Of course, he won't show those results.

I am not trying to hide anything. I already told you that the
performance at high ISO would be improved, and surpass the FZ50 pixels by
a margin a little less than a stop. The 40D would do a little better yet
(and the 450D, a hair better than that, AFAIK, but I haven't enough
experience with its RAW data to know its RAW level sensitivity compared
to the others).

Why do I need to prove what I "admit" to, if no one else is challenging
it?

Believe it or not, there is value in having higher SNR in the deep
shadows at low ISO! In fact, if it had been offered as a possibility by
someone else, in another context, most of my critics here would say,
"Yeah, I want cleaner shadows at ISO 100!", but since one possible
solution came from me, in a context that confounds most readers, ISO 100
shadow performance is mocked and trivialized by the peanut gallery.

And make no mistake, the FZ50 pixels will equal or outperform most DSLRs
per unit of area at ISO 1600, too, other than the Nikon D3, the Canon
20D, 30D, 40D, 400D 450D, 1D3, and 1Ds3. It will outperfrom all CCD
sensors, and the Nikon D2x by a wide margin!

--

<>>< ><<> ><<> <>>< ><<> <>>< <>>< ><<>
John P Sheehy <JPS(a)no.komm>
><<> <>>< <>>< ><<> <>>< ><<> ><<> <>><