# low light

Discussion in 'Digital Photography' started by ipy2006, Mar 7, 2007.

1. ### aclGuest

On Mar 12, 2:11 am, "Bart van der Wolf" <> wrote:
> "John Sheehy" <> wrote in message
>
> news:Xns98F06D6F99D10jpsnokomm@130.81.64.196...
>
> > "Roger N. Clark (change username to rnclark)" <>
> > wrote
> > innews::

>
> >> The problem is that our eyes plus brain are very good at
> >> picking out patterns, whether that pattern is below random
> >> noise, or embedded in other patterns.

>
> What's worse, we see non-existing patterns (e.g. a triangle in the
> following link) because we want to:
> <http://www.xs4all.nl/~bvdwolf/temp/Triangle-or-not.gif>.
>
> > Yes, that is a problem, and that is exactly why you can't evaluate
> > noise by standard deviation alone.

>
> That depends what one wants to evaluate. Standard deviation (together
> with mean) only tells something about pixel to pixel (or sensel to
> sensel) performance. It doesn't allow to make valid judgements about
> anything larger.

As a matter of fact, they don't tell you anything (literally) about
pixel to pixel behaviour. If I tell you that a signal has mean zero
and given standard dev, what else can you tell me about it? Nothing.
It could be anything from an otherwise random time series to a sine
wave to a series of square waves to anything else. It's like knowing
the first two coefficients in an infinite power series (well that's
exactly what it is: the first two coefficients in an infinite power
series).

the reason people use the first two moments (mean and std) is that the
noises under consideration are often assumed to be gaussian, in which
case these 2 qtys completely characterise the noise. this is usually a
good approximation when the noise comes from many different sources.

> Banding could be either calibrated out of the larger
> structure, or an analysis of systematic noise should be done (and care
> should be taken to not mistake Raw-converter effects for camera or
> sensor array effects).

acl, Mar 11, 2007

2. ### Roger N. Clark (change username to rnclark)Guest

John Sheehy wrote:
> "Roger N. Clark (change username to rnclark)" <> wrote in
> news::
>
>> I too agree that pattern noise is more obvious that random noise.
>> Probably by at least a factor of ten. It is our eye+brain's
>> ability to pick out a pattern in the presence of a lot
>> of random noise that makes us able to detect many things
>> in everyday life. It probably developed as a necessary
>> thing for survival. But then it becomes a problem when we try
>> and make something artificial and we see the defects in it.
>> It gives the makers of camera gear quite a challenge.

>
> How does that co-exist with your conclusion that current cameras are
> limited by shot noise?
>
> Saying that current cameras are limited by shot noise means that all future
> improvements lie purely in well depth, quantum efficiency, fill factor, and
> sensor size (you'd probably add "large pixels", but I'd disagree). The
> fact is, a 10:1 S:N on the 1DmkII at ISO 100 would be 1.5 stops further
> below saturation, and 1:1 would be 4.3 stop further below it, if there were
>
> http://www.pbase.com/jps_photo/image/75392571
>
> and that is only statistically, without consideration for the pattern noise
> effects, which widen the visual gap even further.
>

Nice plot. If you look at my past posts, you would also see that
I've said for at least a couple of years 14-bit or higher A/D are
needed too because current DSLRs are limited by 12-bit converters.
Some attacked me in this NG with the idea that "if more than 12-bits were
really needed, then why haven't camera manufacturers done it?"
We'll we now see they have, and I'm sure 14 or more-bits will become a
new standard in future DSLRs.

Regarding fixed pattern noise versus photon Poisson noise, your plot
and some simple illustrations show what is dominant. First clue,
look at the thousands of images on the net. How many show fixed
pattern noise? It is very rare. You tend to see fixed pattern noise
at the very lowest lows in an image. Second, if fixed pattern noise
is really a factor, guess what, you can calibrate most of it out with dark
frame subtraction. I think good examples of fixed pattern noise is
illustrated at:
http://www.clarkvision.com/photoinfo/night.and.low.light.photography
Figure 1, for example shows two merged low light images and fixed pattern
noise is not apparent, nor is it the dominant noise source in the image.
Figure 2 shows the black sky above the Sydney opera house in an ISO 100
20 second exposure. Fixed pattern noise is a little over 1 bit out of 12
in the raw data. It simply is not a factor. But where the scene has
signal, e.g. the lit roof, noise is proportional to the square root
of the signal strength, with photon noise up to 18 out of 4095
in the 12-bit raw file. So, over most of the range, photon noise
dominates. The low end, the bottom few values or bottom couple of bits,
is a combination of photon noise, read noise, and fixed pattern noise.
That gives about 10 bits out of 12 with photon noise as the dominant
noise source. Again, if you work at the low end, calibrate out
the majority of fixed pattern noise with dark frames.

Let's work an example.
Let's assume fixed pattern noise is more objectionable by
10 times random noise (this is a reasonable estimate
for me, and I find fixed pattern noise quite objectionable).
But then with processing, e.g. dark frame subtraction, it can
be reduced about 10x, then filtered and reduced more, all with
minimal impact on resolution. Random photon noise in an image
from can only be reduced by pixel averaging, thus reducing spatial
resolution.

Let's use your full well depth, rounding off to 53,000 electrons.
Fixed pattern noise in DSLRs like the 20D and 1D Mark II are between 1 and
2 bits in the A/D at low ISOs. At low signal levels, line-to-line
pattern noise is on the order of 7 electrons in the 1D Mark II, with
low frequency offset of a few tens of electrons (at ISO 100 fixed pattern
noise appears at about the 1-bit level, which is ~13 electrons. The low frequency
fixed pattern noise is entirely eliminated by a dark frame subtraction,
and line-to-line (what you call 1D) is reduced by about 10X with
dark frame subtraction.

So there are multiple conditions. Here is one example:

ISO 100, 1D Mark II, 53,000 electron full signal:

Signal Photon noise Read Noise Fixed-pattern What noise dominates
(elect- stops (electrons) +A/D noise noise Photon, read, or pattern
rons) (electrons) (electrons)

53,000 0 230 17 ~13 Photon
12,250 -4 110 17 ~13 Photon
3,312 -6 57 17 ~13 Photon
828 -8 29 17 ~13 Photon
207 -10 14 17 ~13 all 3 similar
51 -12 7 17 ~13 read + pattern

The above table demonstrate the the sensor has noise dominated by photon
statistics over most of its dynamic range. Each generation
of cameras that comes out pushed the floor where other noise sources in the
electronics show. It is likely we'll see the 1D Mark III push those limits
a stop or two lower. But photon noise remains and is the ultimate
limit.

Here is another test series that illustrates the above conclusions:
Digital Camera Raw Converter Shadow Detail and Image Editor Limitations:
Factors in Getting Shadow Detail in Images

Figure 6 shows areas from +2 to -7.6 stops. But if you look at the different
raw conversions, you'll see widely different results and wildly different
fixed pattern noise. Then look at Figure 16: the camera jpeg looks pretty
clean with less pattern noise than some of the raw conversions.
So when you say you don't believe photon noise versus fixed pattern noise,
understand the effects of converters too.

Roger

Roger N. Clark (change username to rnclark), Mar 12, 2007

3. ### aclGuest

On Mar 12, 2:53 pm, "Roger N. Clark (change username to rnclark)"
<> wrote:

> And that is why people who evaluate sensors do more than simply
> study the standard deviation of one image. To understand noise sources,

Never claimed otherwise! By the way, why don't people study the full
power spectrum of the noise (ie of a blackframe)? That would give
quite a lot of information (it should allow distinguishing between the
white part of the noise and things like banding). And it should not be
too hard (eg with IRIS, split the channels and FT them). And if you do
that to an average of many frames, you'll be studying repeatable noise
only. Is there some particular reason this isn't done by anybody?
>
> The Nikon D50 Digital Camera:
> Sensor Noise, Dynamic Range, and Full Well Analysis
> http://www.clarkvision.com/imagedetail/evaluation-nikon-d50
>

That's quite interesting, why don't you include dark frames from more
cameras? I'd think that this would be quite useful for people
intending to do very low light work.

acl, Mar 12, 2007
4. ### aclGuest

On Mar 12, 12:15 am, "Bart van der Wolf" <> wrote:
> "John Sheehy" <> wrote in message
>
> news:Xns98F06DCDB2811jpsnokomm@130.81.64.196...
>
> > "Roger N. Clark (change username to rnclark)" <>
> > wrote in
> >news::

>
> >> I too agree that pattern noise is more obvious that random noise.
> >> Probably by at least a factor of ten. It is our eye+brain's
> >> ability to pick out a pattern in the presence of a lot
> >> of random noise that makes us able to detect many things
> >> in everyday life. It probably developed as a necessary
> >> thing for survival. But then it becomes a problem when we try
> >> and make something artificial and we see the defects in it.
> >> It gives the makers of camera gear quite a challenge.

>
> > How does that co-exist with your conclusion that current cameras are
> > limited by shot noise?

>
> Shot noise is a physical limitation, not a man made one. The man made
> limitations can be improved upon.
>

The speed of light is also a physical limitation. Would you therefore
agree to the statement that the top speeds of current spaceships are
limited by the speed of light, and therefore we must work on finding
ways to circumvent that (rather than on finding some better propulsion
system than semi-controlled explosions) ? (I'm not claiming that
banding really is the main limitation, by the way, I actually agree
with Roger and presumably you).

acl, Mar 12, 2007
5. ### Roger N. Clark (change username to rnclark)Guest

acl wrote:
> On Mar 12, 2:11 am, "Bart van der Wolf" <> wrote:
>> "John Sheehy" <> wrote in message
>>
>> news:Xns98F06D6F99D10jpsnokomm@130.81.64.196...
>>
>>> "Roger N. Clark (change username to rnclark)" <>
>>> wrote
>>> innews::
>>>> The problem is that our eyes plus brain are very good at
>>>> picking out patterns, whether that pattern is below random
>>>> noise, or embedded in other patterns.

>> What's worse, we see non-existing patterns (e.g. a triangle in the
>> following link) because we want to:
>> <http://www.xs4all.nl/~bvdwolf/temp/Triangle-or-not.gif>.
>>
>>> Yes, that is a problem, and that is exactly why you can't evaluate
>>> noise by standard deviation alone.

>> That depends what one wants to evaluate. Standard deviation (together
>> with mean) only tells something about pixel to pixel (or sensel to
>> sensel) performance. It doesn't allow to make valid judgements about
>> anything larger.

>
> As a matter of fact, they don't tell you anything (literally) about
> pixel to pixel behaviour. If I tell you that a signal has mean zero
> and given standard dev, what else can you tell me about it? Nothing.
> It could be anything from an otherwise random time series to a sine
> wave to a series of square waves to anything else. It's like knowing
> the first two coefficients in an infinite power series (well that's
> exactly what it is: the first two coefficients in an infinite power
> series).

And that is why people who evaluate sensors do more than simply
study the standard deviation of one image. To understand noise sources,
the standard procedure is to make a series of exposures and analyze
the results from the different test conditions. e.g.:

The Nikon D50 Digital Camera:
Sensor Noise, Dynamic Range, and Full Well Analysis
http://www.clarkvision.com/imagedetail/evaluation-nikon-d50

http://www.clarkvision.com/imagedetail/long-exposure-comparisons/index.html

and more at:
http://www.clarkvision.com/imagedetail/index.html#sensor_analysis

other:
http://www.astrosurf.org/buil/20d/20dvs10d.htm

Roger

> the reason people use the first two moments (mean and std) is that the
> noises under consideration are often assumed to be gaussian, in which
> case these 2 qtys completely characterise the noise. this is usually a
> good approximation when the noise comes from many different sources.
>
>> Banding could be either calibrated out of the larger
>> structure, or an analysis of systematic noise should be done (and care
>> should be taken to not mistake Raw-converter effects for camera or
>> sensor array effects).

>
>

Roger N. Clark (change username to rnclark), Mar 12, 2007
6. ### Roger N. Clark (change username to rnclark)Guest

acl wrote:
> On Mar 12, 2:53 pm, "Roger N. Clark (change username to rnclark)"
> <> wrote:
>
>> And that is why people who evaluate sensors do more than simply
>> study the standard deviation of one image. To understand noise sources,

>
> Never claimed otherwise! By the way, why don't people study the full
> power spectrum of the noise (ie of a blackframe)? That would give
> quite a lot of information (it should allow distinguishing between the
> white part of the noise and things like banding). And it should not be
> too hard (eg with IRIS, split the channels and FT them). And if you do
> that to an average of many frames, you'll be studying repeatable noise
> only. Is there some particular reason this isn't done by anybody?

astrophotography. Then after seeing the trends, it became clear to
me that because the photon noise limit had been reached, one can
model and predict performance pretty closely. Now I find it
interesting about the claims coming out in some press releases
that seem to ignore physical reality ;-).
I and other astrophotographers tend to ignore fixed pattern noise
because we can calibrate most of it out of our images. If that is an
issue for other people, then I suggest they learn how to take
dark frames, average them, and subtract them from their images.
It is really pretty easy, but for best results, it needs to be
done on linear data. Another calibration that can improve images is
flat field calibration, which not only corrects for pixel to pixel
variations, but corrects for light fall-off from lenses.

But if someone wants to pay me to run more tests......

>> The Nikon D50 Digital Camera:
>> Sensor Noise, Dynamic Range, and Full Well Analysis
>> http://www.clarkvision.com/imagedetail/evaluation-nikon-d50

> That's quite interesting, why don't you include dark frames from more
> cameras? I'd think that this would be quite useful for people
> intending to do very low light work.

Again, time. I do have a fair amount of additional data for a number
of cameras but I have not had time to write it up.

Roger

>

Roger N. Clark (change username to rnclark), Mar 12, 2007
7. ### Doug McDonaldGuest

Roger N. Clark (change username to rnclark) wrote:
> acl wrote:
>> On Mar 12, 2:53 pm, "Roger N. Clark (change username to rnclark)"
>> <> wrote:
>>
>>> And that is why people who evaluate sensors do more than simply
>>> study the standard deviation of one image. To understand noise sources,

>>
>> Never claimed otherwise! By the way, why don't people study the full
>> power spectrum of the noise (ie of a blackframe)? That would give
>> quite a lot of information (it should allow distinguishing between the
>> white part of the noise and things like banding). And it should not be
>> too hard (eg with IRIS, split the channels and FT them). And if you do
>> that to an average of many frames, you'll be studying repeatable noise
>> only. Is there some particular reason this isn't done by anybody?

>
> Time and effort--remember most are doing this for free out of
> astrophotography. Then after seeing the trends, it became clear to
> me that because the photon noise limit had been reached, one can
> model and predict performance pretty closely.

I have the Canon 30D. I took a bunch of very underexposed shots
recently (no tripod at critical time) and found that background
subtraction didn't help much. The annoying noise is some sort
of horizontal banding or streaking (these are landscape shots).
Looks sort of like they scan the image TV-wise and this is 1/f noise
in the amplifiers.

Doug McDonald

Doug McDonald, Mar 12, 2007
8. ### Roger N. Clark (change username to rnclark)Guest

Doug McDonald wrote:

> I have the Canon 30D. I took a bunch of very underexposed shots
> recently (no tripod at critical time) and found that background
> subtraction didn't help much. The annoying noise is some sort
> of horizontal banding or streaking (these are landscape shots).
> Looks sort of like they scan the image TV-wise and this is 1/f noise
> in the amplifiers.
>

Doug,
Did you record the raw data, or just jpegs?
You need to record the dark frames under as close to the
same temperature as you can. With the lens cap on
(a dark or dimly lit room is fine too) set the
exposure time to the same as the exposures with the problem.
Record ten to twenty of the. If raw, convert them
with the same settings as your landscape image.
Average all the darks into one master dark, then
subtract the master dark from the landscape frames.
The closer the environmental conditions are to the
landscape images, the better the correction will be.

Roger

Roger N. Clark (change username to rnclark), Mar 13, 2007
9. ### John SheehyGuest

Doug McDonald <mcdonald@SnPoAM_scs.uiuc.edu> wrote in
news:et3rqs\$m2h\$:

> I have the Canon 30D. I took a bunch of very underexposed shots
> recently (no tripod at critical time) and found that background
> subtraction didn't help much. The annoying noise is some sort
> of horizontal banding or streaking (these are landscape shots).
> Looks sort of like they scan the image TV-wise and this is 1/f noise
> in the amplifiers.
>

That's pretty typical of digital cameras in general; it is simply more
visible in cameras with a certain ratio of banding noise to total noise.
For the 30D it should be the same as the 20D (ignoring the 30D's fake,
extra ISOs):

http://www.pbase.com/jps_photo/image/65737967/original

The yellow line represents standard deviation of a blackframe, divided by
10 to fit in with the horizontal and vertical banding noises (they'd be
flat if the entire chart scaled for the the yellow line).

A few things become very clear here; the banding is generally only about
1/10 the strength of the total noise, and yet it is highly visible. With
more read noise, the banding would be less obvious (although it may still
contribute somewhat to visible noise, just without the obvious pattern).
The higher ISOs are all normalized for ISO 100; IOW the values for ISO
200 are divided by two, ISO 400 values are divided by 4, etc, so these
are proportional to electrons as units of noise. All noises decrease as
you get to the higher ISOs, and the total noise looks like it is leveling
off a bit from 800 to 1600, but still had room to improve a little at
3200, but 3200 is "fake" and is really ISO 1600 amplification, multiplied
by two, so it is exactly the same as ISO 1600. The horizontal banding is
still dropping dramatically from 800 to 1600, and seems to have the
capability of dropping even further if the amplification went to 3200 or
even 6400.

--

<>>< ><<> ><<> <>>< ><<> <>>< <>>< ><<>
John P Sheehy <>
><<> <>>< <>>< ><<> <>>< ><<> ><<> <>><

John Sheehy, Mar 13, 2007
10. ### C J CampbellGuest

On 2007-03-07 04:03:00 -0800, "ipy2006" <> said:

> I have to shoot action photos in low light conditions. What is the
> best DSLR for this purpose?
> Thanks,
> Yip

For your budget, any of the cameras that handle well -- a D40 or Rebel
will work nicely, especially if you get a fast 50mm lens.

If you want to use flash, get a real flash unit.
--
World Famous Flight Instructor

C J Campbell, Mar 13, 2007
11. ### John SheehyGuest

"Roger N. Clark (change username to rnclark)" <> wrote
in news::

> I and other astrophotographers tend to ignore fixed pattern noise
> because we can calibrate most of it out of our images.

I'm not sure where "fixed pattern noise" came into play here; the issue
was read noise and one of it's components, 1-D noise. There is, for all
intents and purposes, zero fixed pattern noise in my 20D. Subtracting a
stack of black frames from a short exposure results in nothing but
slightly higher noise.

> If that is an
> issue for other people, then I suggest they learn how to take
> dark frames, average them, and subtract them from their images.

> It is really pretty easy, but for best results, it needs to be
> done on linear data.

And in the case of Canons which have "negative noise" at the blackpoint,
it needs to be done without any clipping at the black level.

> Another calibration that can improve images is
> flat field calibration, which not only corrects for pixel to pixel
> variations, but corrects for light fall-off from lenses.

> But if someone wants to pay me to run more tests......

I don't feel like financing anything right now, but I might suggest that
when you have the time, you do a "gap" test of large vs small pixels.
Your 1DmkII vs S70 page seems to be about pixel size, but it is really
about sensor size. Do a test with a small-pixel camera, and the 1DmkII,
both using the same real focal length, the same Av value, the same Tv
value, the same ISO setting, of the same detailed subject from the same
distance. I guarantee that your big pixels will fall to the ground like
Goliath, when viewing the subject at any magnification, from both
downsampled to both upsampled, or printed large. This is the real test
of pixel size. What you seem to overlook in your analyses is the fact
that standard deviation is only *one* factor in the noise equation;
magnification is another, and the low noise of big pixels is visually
magnified when the pixels are magnified along with the subject.

I am quite certain that the only benefits of big pixels are:

1) quicker readout time and less storage requirements, and

2) slight benefit in photons collection rate per unit of sensor area due
to less wasted space on the sensor (not always realized, however; my
1.97u FZ50, for example, collects about the same number of photons per
unit of area as the 1DmkII, at RAW saturation for the same ISO).

Here is one of my tests; it needs to be redone, because I realized after
doing it that ISO 1600 on the FZ50 is crippled by a very bad amplifier,
that is worse than pushing 100 to 1600. Here is the original, however:

http://www.pbase.com/jps_photo/image/74020772

Don't forget that the 10D images would need to be sharpened more,
sharpening the noise as well.

--

<>>< ><<> ><<> <>>< ><<> <>>< <>>< ><<>
John P Sheehy <>
><<> <>>< <>>< ><<> <>>< ><<> ><<> <>><

John Sheehy, Mar 15, 2007
12. ### aclGuest

On Mar 15, 2:31 pm, John Sheehy <> wrote:
> that standard deviation is only *one* factor in the noise equation;
> magnification is another, and the low noise of big pixels is visually
> magnified when the pixels are magnified along with the subject.

Exactly, and if you don't need the extra pixels you can bin.

>
> I am quite certain that the only benefits of big pixels are:
>
> 1) quicker readout time and less storage requirements, and
>
> 2) slight benefit in photons collection rate per unit of sensor area due
> to less wasted space on the sensor (not always realized, however; my
> 1.97u FZ50, for example, collects about the same number of photons per
> unit of area as the 1DmkII, at RAW saturation for the same ISO).

Well, as long as there are no constant noise sources (eg 10 electrons/
pixel independent of the area). I have no idea if there are or not.

acl, Mar 15, 2007
13. ### PatGuest

On Mar 7, 7:03 am, "ipy2006" <> wrote:
> I have to shoot action photos in low light conditions. What is the
> best DSLR for this purpose?
> Thanks,
> Yip

At the risk of pissing off all of the "purists" out there, you might
want to consider the original Canon Digital Rebel (the good old 300).
That would get you a useable body for not much money. Then add the
Russian operating system to get to ISO of 3200. It's a bit grainy but
sometimes grainy is better than nothing.

Then, with your "extra" money get a Canon 580 flash (or two) and a
"wedding bracket" to avoid red eye and limit shadow. Skip the kit
lens and get the Tokina F2 (or f2.8) zoom. it is something like a 28
to 70mm.

That would get you a servicable package within you price range.

There are lots of situation where this wouldn't be the right setup,
but for what you are describing it will work just fine.

Good luck with it.

Pat, Mar 15, 2007
14. ### Roger N. Clark (change username to rnclark)Guest

John Sheehy wrote:
> "Roger N. Clark (change username to rnclark)" <> wrote
> in news::
>
>>I and other astrophotographers tend to ignore fixed pattern noise
>>because we can calibrate most of it out of our images.

>
> I'm not sure where "fixed pattern noise" came into play here; the issue
> was read noise and one of it's components, 1-D noise. There is, for all
> intents and purposes, zero fixed pattern noise in my 20D. Subtracting a
> stack of black frames from a short exposure results in nothing but
> slightly higher noise.

Fixed pattern noise occurs in different ways with different
sensors. All sensors have fixed pattern noise, even your
20D unless you have a magical one. For example, see:
http://www.astrosurf.org/buil/5d/test.htm
It is in French, but the pictures are labeled well enough
with 30D, 5D etc, that you can see the effects. Common
is vertical striping and amplifier glow. There is no camera,
CCD or CMOS that doesn't have fixed pattern noise.

http://www.clarkvision.com/photoinfo/night.and.low.light.photography
shows that the Canon 1D Mark II has a low level background offset.
That too is fixed pattern noise. So is the line striping
you see in the images on this page. All cameras have these
effects.

A good example of amplifier glow creating an offset near the
edge of the frame is at:
http://www.clarkvision.com/imagedetail/long-exposure-comparisons
e.g. see Figure 2b.

>> If that is an
>>issue for other people, then I suggest they learn how to take
>>dark frames, average them, and subtract them from their images.

>

other signals, regardless of exposure. It is a property of
reading the sensor, not a property of the exposure time.
Examples on the above two web pages show read noise in both
short and long exposures.

>>It is really pretty easy, but for best results, it needs to be
>>done on linear data.

>
> And in the case of Canons which have "negative noise" at the blackpoint,
> it needs to be done without any clipping at the black level.

Sensors collect photons, which are converted to electrons.
The signal is always positive or zero, not negative.
so that the signals do not go negative. Of course,
if noise is too high, then the output signal could hit
zero. Very few pixels are zero in most cameras, even
at the shortest exposure times in the dark.
(I know you know this; I'm adding information to provide
offense; I know you have studied sensors in detail and you have
provided great information to us for years.)
So, I don't know what you mean by negative noise.

>>Another calibration that can improve images is
>>flat field calibration, which not only corrects for pixel to pixel
>>variations, but corrects for light fall-off from lenses.
>>But if someone wants to pay me to run more tests......

>
> I don't feel like financing anything right now, but I might suggest that
> when you have the time, you do a "gap" test of large vs small pixels.
> Your 1DmkII vs S70 page seems to be about pixel size, but it is really
> about sensor size. Do a test with a small-pixel camera, and the 1DmkII,
> both using the same real focal length, the same Av value, the same Tv
> value, the same ISO setting, of the same detailed subject from the same
> distance. I guarantee that your big pixels will fall to the ground like
> Goliath, when viewing the subject at any magnification, from both
> downsampled to both upsampled, or printed large. This is the real test
> of pixel size. What you seem to overlook in your analyses is the fact
> that standard deviation is only *one* factor in the noise equation;
> magnification is another, and the low noise of big pixels is visually
> magnified when the pixels are magnified along with the subject.
>
> I am quite certain that the only benefits of big pixels are:
>
> 1) quicker readout time and less storage requirements, and
>
> 2) slight benefit in photons collection rate per unit of sensor area due
> to less wasted space on the sensor (not always realized, however; my
> 1.97u FZ50, for example, collects about the same number of photons per
> unit of area as the 1DmkII, at RAW saturation for the same ISO).

Here is the fundamental fallacy of your assertion that the
only benefit is better fill factor (that is what you describe in
#2 above): The physics of lenses, and not directly related
to sensors at all.

Every lens at a given f/ratio delivers, for a given light source,
the same surface brightness in th4e focal plane. Another
way to put this is the photons per square micron is constant
at a given f/ratio regardless of the lens focal length.
So an f/4 lens of 20mm focal length looking at a gray
card in sunlight delivers the same number of photons per
square micron to its focal plane as does a 500 mm f/4 lens
looking at the same gray card. It is a simple deduction,
that given two sensors, identical in every way including
quantum efficiency, read noise, and fill factor, that
the sensor with larger pixels collects more photons
simply due to lens physics.

An 8 micron pixel collects 16 times the photons as a
pixel 2 microns in size (8*8/(2*2) = 16), and that is exactly
what we observe with today's digital cameras. For example,
see:
Digital Cameras: Does Pixel Size Matter?
Part 2: Example Images using Different Pixel Sizes
http://www.clarkvision.com/imagedetail/does.pixel.size.matter2

> Here is one of my tests; it needs to be redone, because I realized after
> doing it that ISO 1600 on the FZ50 is crippled by a very bad amplifier,
> that is worse than pushing 100 to 1600. Here is the original, however:
>
> http://www.pbase.com/jps_photo/image/74020772
>
> Don't forget that the 10D images would need to be sharpened more,
> sharpening the noise as well.

Your test is biased in that the two images from the two cameras
are not comparable. By using two different sized sensors
with the same focal length, of course the sensor with
smaller pixels sees finer detail. But the large sensor
shows a larger field of view that is not covered by the
smaller sensor at all. So depending on who wanted the
image, one could draw different conclusions: the person who
wanted a wide field of view would choose the large sensor;
one who wanted a telephoto image would choose the small
pixels. But in either case, the pixels from the small
sensor would be noisier in proportion to the square root
ratio of the areas of each pixel.

Roger

Roger N. Clark (change username to rnclark), Mar 16, 2007
15. ### David J. LittleboyGuest

"Roger N. Clark (change username to rnclark)" <> wrote:
> John Sheehy wrote:
>>
>> I don't feel like financing anything right now, but I might suggest that
>> when you have the time, you do a "gap" test of large vs small pixels.
>> Your 1DmkII vs S70 page seems to be about pixel size, but it is really
>> about sensor size. Do a test with a small-pixel camera, and the 1DmkII,
>> both using the same real focal length, the same Av value, the same Tv
>> value, the same ISO setting, of the same detailed subject from the same
>> distance. I guarantee that your big pixels will fall to the ground like
>> Goliath, when viewing the subject at any magnification, from both
>> downsampled to both upsampled, or printed large. This is the real test
>> of pixel size. What you seem to overlook in your analyses is the fact
>> that standard deviation is only *one* factor in the noise equation;
>> magnification is another, and the low noise of big pixels is visually
>> magnified when the pixels are magnified along with the subject.
>>
>> I am quite certain that the only benefits of big pixels are:
>>
>> 1) quicker readout time and less storage requirements, and
>>
>> 2) slight benefit in photons collection rate per unit of sensor area due
>> to less wasted space on the sensor (not always realized, however; my
>> 1.97u FZ50, for example, collects about the same number of photons per
>> unit of area as the 1DmkII, at RAW saturation for the same ISO).

>
> Here is the fundamental fallacy of your assertion that the
> only benefit is better fill factor (that is what you describe in
> #2 above): The physics of lenses, and not directly related
> to sensors at all.
>
> Every lens at a given f/ratio delivers, for a given light source,
> the same surface brightness in th4e focal plane. Another
> way to put this is the photons per square micron is constant
> at a given f/ratio regardless of the lens focal length.
> So an f/4 lens of 20mm focal length looking at a gray
> card in sunlight delivers the same number of photons per
> square micron to its focal plane as does a 500 mm f/4 lens
> looking at the same gray card. It is a simple deduction,
> that given two sensors, identical in every way including
> quantum efficiency, read noise, and fill factor, that
> the sensor with larger pixels collects more photons
> simply due to lens physics.

I think you guys are talking past each other here.

I think John is arguing that _for a sensor of a given size_, larger pixels
aren't any better.

David J. Littleboy
Tokyo, Japan

David J. Littleboy, Mar 16, 2007
16. ### Roger N. Clark (change username to rnclark)Guest

David J. Littleboy wrote:

> I think you guys are talking past each other here.
>
> I think John is arguing that _for a sensor of a given size_, larger pixels
> aren't any better.

1) Well, his example used 2 different sized sensors.

2) There is a difference. The signal you record has added
read noise. A larger pixel collects more photons
so the signal is larger compared to the read noise.
Thus you can detect fainter things, or have better high
ISO performance. If you sum the signal from a smaller
pixels to equal the area of a larger pixel size,
as much as having the larger pixel with one read noise.

Roger

Roger N. Clark (change username to rnclark), Mar 16, 2007
17. ### LionelGuest

On Fri, 16 Mar 2007 12:29:52 +0900, "David J. Littleboy"
<> wrote:

>I think you guys are talking past each other here.
>
>I think John is arguing that _for a sensor of a given size_, larger pixels
>aren't any better.

But they /are/ better! - That's why the sensor designers are
constantly trying to improve the fill-factor, ie; make the pixels (or,
more accurately, the actual photo diode surface, which is smaller than
the pixel size) bigger for a given sensor size/resolution ratio. This
is because the bigger the suface of the photodiode (as a proportion of
the size of that pixel on the sensor), the more photons it'll collect
for a given exposure. And, all else being equal, more photons equals a
better signal to noise ratio.

--
W "Some people are alive only because it is illegal to kill them."
. | ,. w ,
\|/ \|/ Perna condita delenda est
---^----^---------------------------------------------------------------

Lionel, Mar 16, 2007
18. ### aclGuest

On Mar 16, 6:29 am, "David J. Littleboy" <> wrote:

> I think you guys are talking past each other here.
>
> I think John is arguing that _for a sensor of a given size_, larger pixels
> aren't any better.
>

But doesn't this make him a "crop fan" for you? Or does your attitude
depend on who you're replying to?

acl, Mar 16, 2007
19. ### aclGuest

On Mar 16, 7:04 am, "Roger N. Clark (change username to rnclark)"
<> wrote:

> 2) There is a difference. The signal you record has added
> read noise. A larger pixel collects more photons
> so the signal is larger compared to the read noise.
> Thus you can detect fainter things, or have better high
> ISO performance. If you sum the signal from a smaller
> pixels to equal the area of a larger pixel size,
> as much as having the larger pixel with one read noise.
>

Is read noise fixed per pixel, per unit area, or something else?

acl, Mar 16, 2007
20. ### Roger N. Clark (change username to rnclark)Guest

acl wrote:
> On Mar 16, 7:04 am, "Roger N. Clark (change username to rnclark)"
> <> wrote:
>
>> 2) There is a difference. The signal you record has added
>> read noise. A larger pixel collects more photons
>> so the signal is larger compared to the read noise.
>> Thus you can detect fainter things, or have better high
>> ISO performance. If you sum the signal from a smaller
>> pixels to equal the area of a larger pixel size,
>> as much as having the larger pixel with one read noise.
>>

>
> Is read noise fixed per pixel, per unit area, or something else?

Read noise is per pixel. Say you had 2 sensors, one with half
the pixel size, so you needed to add 4 pixels to equal the area
of the larger pixel. Lats say both had great read noise of
4 electrons. The larger pixel gets: X + 4 electrons noise.
The smaller pixel sensor, adding 4 pixels gets:
X + sqrt(4)*4 = X + 8, so the read noise is effectively
doubled.

Read noise for a given sensor is dependent on the design of the sensor
and how the readout is configured. Read ranges from just under 4
to about 30 electrons and is not dependent on pixel size.
For example, see Figure 3 at;
http://www.clarkvision.com/imagedetail/digital.sensor.performance.summary

At low ISO, and low bit count (e.g. 12 bits) and noise in the A/D converter
contributes greater noise than the true read noise from the sensor.

Roger

Roger N. Clark (change username to rnclark), Mar 16, 2007