Re: D3x great camera, but...
- From: "Roger N. Clark (change username to rnclark)" <username@xxxxxxxxx>
- Date: Sun, 01 Feb 2009 11:54:26 -0700
John,
If you would like an intelligent open discussion about this
topic I am open, even as private emails, but it is not
productive to launch attacks in a public forum.
I'm a little surprised by some of the statements you made
below because I do not remember some of the allegations
(but then I'm jet lagged having just returned from Tanzania).
As someone has alerted me to this thread, I'll address a few
issues, but I'll not get involved in bashing.
For everyone: John has some good points and I am pretty sure
I have said this before. Where I disagree with John is in
considering the whole picture (pun not intended).
I'll address a couple of specific points raised then
look at the general problem of pixel density that John
raises.
John Sheehy wrote:
nospam <nospam@xxxxxxxxxxxxxx> wrote in
news:250120090053095973%nospam@xxxxxxxxxxxxxx:
In article <Xns9B9DE78A0811Ajpsnokomm@xxxxxxxxxxxx>, John Sheehy
<JPS@xxxxxxx> wrote:
I can't believe that there are so many seemingly college-educated
people who can't keep their heads straight in these discussions. My
statement is that the best P&S sensors capture more photons *PER UNIT
OF AREA* with the same exposure, than the best DSLRs. Why is that so
difficult for so many people to understand?probably because it's incorrect.
In theory, if the quantum efficiency were constant, as you
decrease pixel size, its like cutting a pie into smaller pieces, so
there would be no change in light sensitivity per unit area
as there is no change in the amount of pie (for the pie analogy).
However, when one considers the real world, a pixel must have
an active area and an inactive area, a wall so to speak to
keep the converted electrons in the pixel. Without the wall,
electrons will leak into adjacent pixels, known as bleed. So as
pixel size decreases, the dead space becomes a larger percentage
of the area and fill fraction drops. And this is ignoring other
issues like in CMOS needing the support electronics in each pixel,
which only makes the problem worse. (The support electronics
on CMOS sensors, by the way, is a major reason small pixel size
sensors use CCDs, as the fill factor can be larger.)
Then there is the "ugly fact" that the basic principle of photon
absorption in silicon sets the well size. The 1/e absorption length
in silicon is about 1 micron for blue light, 3 microns for green
light, 6 microns for red light, and about 8 microns for the upper
end of the red filter cut-off. That means that red light,
for example would take several 2-micron pixel lengths to be absorbed
(not good). This is masked to a degree in real cameras by the
Bayer sensor design and the need for a blur filter.
The resulting effects are on the whole toward less efficiency with
smaller pixels. The actual effects on any one sensor can be large,
so one sensor may trade one problem for another and make it look
a little better in one respect, e.g. by changes to voltage biases.
John, your comparison of an FZ50 CCD versus 400D CMOS sensor adds
complications to your conclusions, especially regarding smaller versus
larger pixel sensitivities.
It is not incorrect. Every measurement done correctly, and every truly equal comparison I've seen, shows that P&S-sized sensors are as good or better than DSLR sensors at photon efficiency (NOT TOTAL QUANTITY!), and that base ISO read noise, integrated by pixel pitch, is also equal or lower (at least before the D3x came along - which has the highest pixel density of DSLRs; not one of the lowest).
John, How many measurements are there? Could you please cite them
and give links? This is an honest question as I would like to add them
to my sensor performance summary page. And these are Quantum
Efficiency measurements? QE is the real measure of sensitivity,
or do you really mean something else?
How do you account for the fact that Emil and I, independently, measured that the FZ50 collects 4800 maximum photons per photosite, and that this occurs at 2.5 stops above ISO 100 middle grey?
What did Roger measure? Nothing. He said it was impossible because he didn't think it could be that smaller pixels were more efficient. A priori reasoning all the way.
Where did I say it was impossible? If you look at Figure 1 at:
http://www.clarkvision.com/imagedetail/digital.sensor.performance.summary
you will see my model indicates that an FZ50-sized sensor
(1.97 micron pixel pitch) will produce a similar full well capacity.
In fact the electron density
(4800 e- / 1.97^2) = 1237 electrons/ sq. micron
is on the low side.
Typical ranges are 1000 to a little over 2000 electrons
per sq micron for silicon based sensors.
The 5D mark 2 is 1600 electrons/sq micron for example.
This ignores the dead space. If we assume a 0.5 micron
dead space the densities would be:
FZ50: (4800/(1.97-.5)^2) = 2221 e-/sq micron
5D2: (65700/(6.4 -.5)^2) = 1887 e-/sq micron
But neither full well capacity nor electron densities are
indicators of sensitivity. You need quantum efficiency.
Note: higher electron densities can be detrimental to image quality,
but in either case, these two are in the same ballpark and not out
of the ordinary. But for total pixel area, the FZ50 is showing the
fill factor is smaller and the efficiency of light gathering
is lower. Quantum efficiency is a different matter.
The fact is, Roger's analyses have ignored the real sensitivity of cameras, assuming that the maximum number of photons at any ISO in a P&S occurs at the same relationship to middle grey as it does in DSLRs,
Excuse me, I report photoelectrons that are detected. My site is about
sensor performance on an equal footing, thus I treat all sensors with
the same standards so they may be best compared. How a specific
manufacturer sets their gains is irrelevant and if the user knows
the sensor characteristics they can properly compensate.
The "REAL" sensitivity of cameras is the system throughput, which
involves the Quantum Efficiency and transmission of the optics
over the sensor (Bayer color filter array, blur and IR filters).
To measure this one needs standard light sources and/or a reference
of known quantum efficiency. Few people have done this (I have done
a couple of cameras, and Christian Buil just put up a page with some more).
I think you are confusing manufacturer A/D output with true
sensitivity.
I can't believe that there are so many seemingly college-educated people
who can't keep their heads straight in these discussions. My statement is
that the best P&S sensors capture more photons *PER UNIT OF AREA* with the
same exposure, than the best DSLRs. Why is that so difficult for so many
people to understand? Why does almost everyone turn their minds to things
like Roger's silly, rhetorical S60 vs 1Dmk2 noise comparison? Those are
two *COMPLETELY DIFFERENT SIZE SENSORS*. A crop from the 1Dmk2, the
physical size of an s60 sensor, would have far less resolution, less shot
noise, and less read noise except at ISO 800 and above. All done with
smaller pixels.
John, it is not silly, it is the real world as it exists today.
People want to know the difference between choosing a large sensor
and small sensor, and that demonstration shows the effects.
What you are advocating is a completely different case. It does not
make my comparison silly. Do you understand that?
When the consumer choice is between a 20 megapixel camera with
large pixels and a 200 megapixel camera with the same sized
sensor, I'll do that comparison, but it doesn't exist now.
John, your comparison at:
http://www.pbase.com/jps_photo/image/100825882
is interesting, but incomplete. The 40D trades spatial resolution
for lower noise. This gets into subjective imaging, like
in the film versus digital: some liked the grain of film while
others liked the smoothness of digital. I'll wager the
information content in the two images is very similar.
Would you email me the 40D frame? I'll
run Richardson-Lucy deconvolution on it and trade some
noise for improved spatial resolution. I bet there would be
little difference between that and FZ50 image. I'll email
it back to you and you can post the comparison. The point is that
I think the images are actually more similar in total information
content than it appears.
But try this same experiment with true low light work, as in
detecting a faint galaxy. It will show the increased total
read noise of the FZ50 will not be able to detect stars
or nebula as faint. There is a simple mathematical
demonstration of this.
Consider 2 sensors of the same size, one with A) 2-micron pixels,
and one with B) 8 micron pixels. Assume equal QE and 100% fill factors.
Assume read noise on both sensors is 2 electrons. There are
16 two-micron pixels for each 8-micron pixel. Consider an
exposure that captures 1 photon/square micron. The signal-to-
noise ratios on each camera per 64-square microns is
(S = signal, N = total noise, R = read noise = 2):
A) S = 64, N= sqrt(S + 16*R*R) = 11.3, S/N = 5.66
B) S = 64, N= sqrt(S + 1*R*R) = 8.25, S/N = 7.76.
Noise per pixel:
A) S = 4, N= sqrt(S + 1*R*R) = 2.83, S/N = 1.41
B) S = 64, N= sqrt(S + 1*R*R) = 8.25, S/N = 7.76
You see, you can't win unless read noise is truly zero.
> Shot noise is not a major issue in current DSLRs. The problems people
> are having with noise in their DSLR photography, especially full-frame,
> are almost exclusively read noise.
Another thing John, you seem to often say that read noise is
the most significant noise factor in images. Take some typical
images and compute the noise sources at each level in the
image. Here is a typical image, for example:
http://www.clarkvision.com/galleries/gallery.bird/web/lilac-breasted.roller.c01.24.2007.JZ3F1238.b-700.html
Where in that image do you think read noise dominates?
How about this one where noise is more prominent?
http://www.clarkvision.com/galleries/gallery.africa/web/vulture.in.fog.c01.20.2007.JZ3F8185b-700.html
Or a faint galaxy:
http://www.clarkvision.com/galleries/gallery.astrophoto-1/web/m33-c07.10.2005-AdpAdd-v1.4-700.html
I can tell you in each of these images, photon noise dominates
over read noise in all areas except the dark sky in the galaxy
image and the darkest shadows in the lilac breasted roller bird
image. Is anyone bothered by the noise in the shadows of the
lilac breasted roller image?
The vulture image has noise entirely dominated by photon noise.
So John, I disagree that read noise is a significant factor in
most digital photographs. The dominant noise source is photon
noise. The second most annoying noise source is pattern noise.
Read noise is last.
=====================
OK, so now let's look at the general problem that John raises.
The sensor in the focal plane is sampling the image. The finer
the samples the better the possible resolution. So John's idea
is for the SAME SIZED SENSOR to have more pixels to provide better
resolution is sound to first order. And that improvement in
resolution with smaller pixels has been what we have been experiencing
with the series of cameras like the Canon 10D, 20d/30D, 40D, and 50D
going from 6 to 8 to 10 to 15 megapixels with the same APS-C sized
sensor. Is the 60D going to be 25 or 30 megapixels, 70D 50 megapixels,
etc?
People are complaining about the noise in both the 50D and G10.
So why aren't these cameras better with their higher pixel densities?
Well, actually they do produce some stunning images in good light.
But problems show in lower light/high ISO (yes, this is partly due
to fixed pattern noise/banding).
Let's consider a perfect sensor: 100% quantum efficiency, 0 read
noise, zero fixed pattern noise. There actually are sensors
close to these ideals, such as electron multiplier CCDs that
multiply the electron count making read noise irrelevant. They
are used in Astronomy and military surveillance. But they suffer
from some undesirable qualities for general imaging: limited dynamic
range. For example, with 15x electron multiplication, the full well
is 15x smaller (referring to photons detected).
And this is the case for smaller sensors: smaller full well capacity
and smaller dynamic range. Let's take the extreme: pixels so small
the full well is 1 electron. So now we have an image of only 1's and
0's. So to get any tonality, one must average over a large area.
This clearly indicates a lower limit. (Film with it's grain acts
differently so is not comparable, and one must average over many
film grains too to get an image with nice tonality.)
So the small sensors have limited dynamic range and because that
property is dictated by the full well capacity (even in a zero read
noise sensor) and that property is fundamentally limited by the
absorption lengths of photons in silicon, it is not likely to
change. To get greater full well capacities, the absorption
lengths must be increased, just the opposite of what one needs
with smaller pixels.
On the low end, with zero read noise, one in theory could average/add
pixels together to get any sized pixel you wanted to increase the
signal. This would certainly be an ultimate in detection ability
and I agree with John on this aspect. It would not matter what the
pixel size would be. Then you would be able to trade spatial
resolution for signal-to-noise in post processing.
The problems in the real world are several.
1) as noted above, the need for "walls" between pixels to
keep electrons from leaking, affecting efficiency.
2) Photons are finite. For a properly exposed scene in your
digital camera, on an 18% gray card, there are about 1300
photons per square micron incident in the focal plane integrated
over the green passband (this is before any color, blur, or IR
filters, or the loss from less than 100% QE). In full sunlight
there in only about 13,000 photons /second incident in the focal
plane per square micron integrated over the green passband (you can
check my math--remember I'm jet lagged) at f/16 from an 18% gray card.
So you want to expose at ISO 1,000,000? At ISO 1,000,000 in full
sunlight that would be 1/1000000 second exposure at f/16 or
0.013 photons per square micron! The point is there are
practical limits (not what the absolute exact numbers are).
3) As pixel density increases, readout time increases for a given
sensor size. The fastest readout rates are on the order of
100 megapixels/second for current high end cameras. A full frame
sensor with 2-micron pixels would be 18000 x 12000 = 216 Megapixels,
so 2 second readout rates for current high end technology.
For sports shooters, that rate would need to increase to
10 gigapixels /second to get to 10 frames/second.
Hmmm... when does speed of light effects come into play?
Even more important are electron propagation speeds.
4) Diffraction.
Diffraction spot sizes are illustrated in Figure 8 at:
http://www.clarkvision.com/imagedetail/does.pixel.size.matter
For example, at 5-micron pixel pitch, the MTF at f/8, green
light is about zero! I agree in super-sampling the diffraction
disk can be desirable in some cases, but as pixel pitch goes smaller,
diffraction effects limit photographic resolution.
5) Smaller pixels have lower dynamic range. This is a basic
fact of smaller well capacity. This means more difficulty in
exposure and rendering the scene. More pixels does not help
much as John once proposed (statistically it can help by a fraction
of a stop--not enough).
So overall, increasing pixel density helps image quality to a point,
but then other factors come into play and total image quality
degrades. What that trade point is will be a matter of debate.
My model for these effects are described in my Apparent Image
Quality (AIQ) model at:
http://www.clarkvision.com/imagedetail/digital.sensor.performance.summary/index.html#AIQ
For diffraction limited performance at f/8, I find the
peak performance at about 5-micron pixel pitch and this is
in the ballpark of where DSLRs are these days.
In order to move the peak to lower pixel pitches, lens
performance must be improved (e.g. diffraction limited
performance at f/4), with higher quantum efficiency. Even
then I don't see the peak performance going below about
3-micron pixel pitch because of too low of full well capacity.
The f/8 optimum peaks for a full size sensor at just over 30 megapixels
in my model. We are close to that in the Nikon D3X and Sony,
and not far off in the 5D2. (I just returned from Tanzania having
photographed with a 5D2 and 1DII and I must say the 5D2 images
are simply stunning.)
So John's basic theory of higher pixel density has merit, but reality
from a number of factors will limit how far the density can go.
Note I did not factor in file sizes as memory and computer speed
will continue to improve (I have also been processing large format
scanned film files for years so do OK with 200 megapixel equivalent
files with existing technology).
Finally, there are practical limits to manufacturing mass
produced sensors. As pixel sizes decrease, fixed pattern problems
must be reduced. If read noise is less than an electron,
and full wells are a couple thousand electrons, think of the
uniformity requirements over tens of millions of pixels.
It works ok for small chips, but as sensor size increases, this
will get more difficult.
I will continue intelligent and civil discussions, but will leave
when the name calling starts. I have better things to do than
bicker.
Roger
.
- Follow-Ups:
- Re: D3x great camera, but...
- From: Paul Furman
- Re: D3x great camera, but...
- Prev by Date: Re: Sorry OT - Needed USB Camera to use with "Timershot"
- Next by Date: Re: DPR Review SLAMMED
- Previous by thread: Re: D3x great camera, but...
- Next by thread: Re: D3x great camera, but...
- Index(es):
Relevant Pages
|
Loading
