What about 1DMkIII's Dynamic Range?

[DSPographer]

There are two main sources of black point noise. One is dark current which is a leakage of electrons in the photodetector in the absence of light which obeys Poisson statistics (it is an electron count) and the other is pre-amplifier noise. The RMS pre-amplifier noise is not at a level of the square root of the detected signal so it is not Poisson distributed; in general this noise is zero mean so it does not contribute to the mean black level measured. The total dark noise is the combination of these two independent noise sources (their variances add). If the black level is 25 e- (RMS noise of 5 e-) and the RMS pre-amplifier noise is 4 e- (I have seen numbers close to this level quoted for Canon's CMOS sensor pre-amplifiers) then the total dark noise is sqrt(5^2 + 4^2) or 6.4 e- RMS. This noise is still much less than the mean black level of 25 e- that is measured.

 

[DSPographer]

If the raw images include many masked photosites then we can compute the mean and standard deviation of the functional dark sites (some bad sites may be mapped out both in the masked and unmasked portions of the detector). The mean black level for a given pre-amplifier (there are 8 pre-amplifiers in the 1DIII) is the black level that the raw converter subtracts and the standard deviation is the RMS dark noise that remains after the subtraction. We should therefore be able to estimate this total dark noise directly as the stdev of the dark photosite signals.

 

[Its RKM]

How does increasing the pre-amplifier gain by setting a high ISO
improve performance if it is not by reducing the effect of the A-D
noise floor?

Can you rephrase that. I don't know what you're asking.

Its a fairly simple question if you understand what is happening.

My confusion is not about what is happening; I don't understand
your question, grammar-wise. Your question is vague, and
incoherent. I'm not telepathic, nor do I like to assume I know
what someone said when they were not perfectly clear.

Let's be clear on this, it isn't MY question.

I ANSWERED because the question that DSPographer asked (ie. it was HIS question) was trivial, rhetoric and easy to understand. There was nothing confusing, vague, incoherent or unclear about it.

Increasing the ISO simply changes the pre-amp gain between the chip
readout and the ADC. The chip has a noise floor, the actual noise
limits of the sensor, and the ADC also has a noise floor - ideally
just quantisation noise but usually some input noise too.

My point is that it is MOSTLY input noise, as you call it, and
bit depth is the lesser of the two issues, regarding DR.

Not usually. Take a look at some ADC specifications and you will find the input noise is typically 1LSB or less. There certainly are ADCs where the input noise exceeds 1LSB, but that is not always the case.

Changing ISO just chages the gain (the amplification) of the chop
output relative to the ADC input - and so changes the relative
contribution of chip noise and ADC noise to the total output noise.

Higher ISO results in chip noise dominating over ADC noise, lower
ISO results in ADC noise diminating chip noise.

You don't know that.

Yes I do! And so would you if you read the relevant reviews and analyses.

For example, take a look at Tables 1a & 1b in http://www.clarkvision.com/imagedetail/evaluation-1d2/index.html

This shows quite clearly that at ISO400 and below the ADC noise (input noise and quanitisation noise) limits the dynamic range. The sensor still outputs the same noise floor, but it is only AFTER that has been amplified sufficiently (around ISO400) that this noise exceeds the ADC noise, hence reducing the dynamic range of the combination of sensor and ADC.

Simple as that.

At low ISO you NEED lower ADC noise, ie. lower quantisation noise
(ie more ADC resolution) to get everything out of your sensor chip
that it can deliver. That is less of an issue at higher ISO, where
the pre-amp gain ensures that ADC noise is dominated by the sensor
noise, so more bits are irrelevant.

But the fact is, the data is not bit-limited. It's limited by
analog noise BEFORE quantization.

Of course it is limited by analogue noise before quantisation - nobody has suggested that the CMOS chip had infinite signal to noise or dynamic range. However the data is further limited by the ADC when ISO is low. Under those conditions it certainly IS "bit-limited" as you put it. This is clearly demonstrated in Roger Clarke's data and charts.

 

[blackhawk13]

... whilst more bits may enhance the DISPLAY [by any means] of the
data, it will NOT increase the number of stops of information
captured. That is a finite amount. You can add a 128 bit A-D
converter to an old 1D if you are clever enough ... but it won't
help the DR.
An 8-bit camera could EASILY have better DR than a 16-bit unit.
KP

--



http://www.ahomls.com/photo.htm
http://www.phillipsphotographer.com
Voted Best of the City 2004 by Cincinnati Magazine
I don't believe in fate, but I do believe in f/8! And while you're
at it, don't be afraid to vote Libertarian, for REAL freedom!

--The ability to capture a smaller value means a higher dynamic range. In a digital image, the number of possible colors or shades of gray that can be included in a particular image

http://www.dpreview.com/learn/?/Glossary/Digital_Imaging/dynamic_range_01.htm

-nothing beats a fast lense, except a fast girl-

 

[Its RKM]

Bit depth could limit DR, if there were not enough noise, but
there's usually plenty of it.

Not according to
http://www.clarkvision.com/imagedetail/evaluation-1d2/index.html

Below ISO 400, ADC quantisation IS the noise limit.

I don't agree with Roger on many things. I'm currently debating
him on usenet.

I don't agree with everything he says either, but I certainly do agree with him on this issue. The derivation of the procedure he uses is straight forward, the process is easy to reproduce and the results can be independently verified.

Furthermore, its exactly in line with what I would expect and I design imaging sensors for a living - albeit not visible band sensors.

One difference between the RC & I is some of the terminology he uses, which sometimes leads to confusion. For example, in this context he refers to "read noise" as the total noise derived from his measurements - ie. the noise from all sources. Normal convention in sensor design is to use the term "read noise" specifically to refer to the excess noise injected by the read circuit. That read noise is then amplified and added in quadrature to the input noise and quantisation noise of the ADC. Table 1b shows quite clearly that when amplified sufficiently, the read noise of the 1D-II sensor, including the input noise of the ISO amplifier, is around 4 electrons, since at high ISO that is always the dominant noise. However, at ISO400 the total noise is about 1.4x higher than the sensor read noise. That excess noise is the ADC noise, and since the total noise is about sqrt(2) higher than the sensor read noise, then is it obvious that the ADC noise is approximately equal to the sensor read noise at ISO400. For lower ISOs the ADC noise dominates, resulting in a total noise increasing in inverse proportion to ISO. (in RC's tables the total noise is referred back to electrons at the storage site on the sensor.)

 

[DSPographer]

...
But the fact is, the data is not bit-limited. It's limited by
analog noise BEFORE quantization.

--
John

Although an A-D converter may intentionally be designed with some analog noise at its input to insure that it is properly dithered if an A-D converter has a noise floor much worse than the number of bits it contains would suggest then it contains bits which are essentially worthless (I would normally expect a real 12 bit converter to have a DR near the range of an ideal 11 bit converter). Although exaggerated bit depths are sometimes seen in marketing blurbs from unscrupulous companies I don't expect that the A-D converters in Canon's d-slrs suffer from it. I would therefore expect that when they increase the bit depth by 2 bits the dynamic range would increase by something near 2 stops. If the analog input noise dominated the 12 bit converter noise and remained the same in a 14 bit converter then Canon would be guilty of wasteful (in terms of raw data storage) and misleading specsmanship for the 1DIII.

 

[David Martin]

...
But the fact is, the data is not bit-limited. It's limited by
analog noise BEFORE quantization.

--
John

Although an A-D converter may intentionally be designed with some
analog noise at its input to insure that it is properly dithered if
an A-D converter has a noise floor much worse than the number of
bits it contains would suggest then it contains bits which are
essentially worthless (I would normally expect a real 12 bit
converter to have a DR near the range of an ideal 11 bit
converter). Although exaggerated bit depths are sometimes seen in
marketing blurbs from unscrupulous companies I don't expect that
the A-D converters in Canon's d-slrs suffer from it. I would
therefore expect that when they increase the bit depth by 2 bits
the dynamic range would increase by something near 2 stops. If the
analog input noise dominated the 12 bit converter noise and
remained the same in a 14 bit converter then Canon would be guilty
of wasteful (in terms of raw data storage) and misleading
specsmanship for the 1DIII.

I don't see why they wouldn't trumpet it, if they thought they had got it.

I thought that more bits could mean smoother transitions, even over the same DR, at least for JPEG?
--
Regards,
DaveMart

'Just a wildebeest on the plain of life'
Please see profile for equipment

 

[DSPographer]

I don't see why they wouldn't trumpet it, if they thought they had
got it.
I thought that more bits could mean smoother transitions, even over
the same DR, at least for JPEG?
--
Regards,
DaveMart

'Just a wildebeest on the plain of life'
Please see profile for equipment

I think they are trumpeting it when they say 14 bit A-D. They are being careful to note that the D-R of the sensor itself (which exceeds the D-R of the A-D) is unchanged.

 

[John Sheehy]

Of course it is limited by analogue noise before quantisation -
nobody has suggested that the CMOS chip had infinite signal to
noise or dynamic range. However the data is further limited by the
ADC when ISO is low. Under those conditions it certainly IS
"bit-limited" as you put it. This is clearly demonstrated in Roger
Clarke's data and charts.

No, it is not. I've never said that the ADC section of the signal chain couldn't be responsible for the flat rate of noise. I believe that it probably has a lot to do with it. What I have said is that it is not the BIT DEPTH , which is a different thing. In other words, what I am saying is that if you did have a 14-bit ADC that had 2 ADU of read noise, and quantized its output to 11 bits, it would have more dynamic range than a 12-bit with 2 ADU of read noise, with the full 12 bits. IT'S THE ANALOG NOISE, NOT THE BITS that limits DR with current cameras.

--
John

 

[DSPographer]

...
I thought that more bits could mean smoother transitions, even over
the same DR, at least for JPEG?
...

If the transitions are stepped because of the A-D converter then the A-D converter is limiting the noise floor performance. If analog amplifier noise and/or dark current shot noise is limiting the noise floor then the A-D converter is sufficiently dithered to avoid stair-stepping or posterization.

 

[Peter Carmichael]

Because the raw capture is a linear count directly related to
incident photons, the dynamic range of the captured image can be
limited by having too few bits available to digitize the available
signal information. 14 bits stores two more stops of sensor data
than 12 bits. Fact. Not fiction. That's just how it is.

I do have a conceptual issue with this fact. To me, more bits on
the sampling "depth" just add more granularity, or increased
"signal" awareness, as I call it. More bits are the "conduit" for
much smoother ramps or "gradients" of tonality, in general, if we
try to illustrate the actual effect / application.

This is tricky to describe. I'll try two approaches to describing it.

1. Dictionary definition

14 bits represent numbers from 0 to 16383. 12 bits represent numbers from 0 to 4095. In both cases the smallest variation is unity (one). The definition of dynamic range is... the ratio between the largest value to the smallest discernible change. 16383:1 vs. 4095:1.

2. Think of the added bits as being the most significant bits...

The 0-4095 range of the 12 bit A/D converter is used to digitize a value varying from 0 - 5 volts. A value of 5 volts is represented by 4095. The smallest discernible change is 5/4095 = 0.001221 volts. The 0-16383 range of the 14 bit A/D converter is used to digitize the same signal. A 5v signal still results in a value of 4095. The smallest discernible variation is still 0.001221 volts. However, the extra two bits allow this converter to handle input values up to 20 volts, represented by 16383.

 

[John Sheehy]

It isn't clear to me where you are going with your argument. About
the black point you always have noise due to the electronic
amplification chain including the front end of the A/D conversion
(i.e. you can't escape noise at or near the black point).

The person I responded to claimed that a blackframe has photon noise or shot noise.

You have to be counting photons to experience any counting noises. A blackframe only has noise from dark current, pixel quirks, and read noises, including any noise added by the ADC.

--
John

 

[DSPographer]

...

You have to be counting photons to experience any counting noises.
A blackframe only has noise from dark current, pixel quirks, and
read noises, including any noise added by the ADC.

No you only need to be counting something for the standard deviation in the count to depend on Poisson statistics. All of the mean black level is due to a count of electrons and therefore it contains shot noise in proportion to the square root of the number of electrons counted. There is additional read noise from the preamplifier but since that doesn't shift the mean dark current measured it's value is not related to mean the dark level. As I pointed out below the unknown component of the dark level will be the dark noise remaining after the raw converter subtracts the dark level and this noise is due to the standard deviation of the dark level.

 

[John Sheehy]

Although an A-D converter may intentionally be designed with some
analog noise at its input to insure that it is properly dithered

That might be necessary with 8-bit audio (and 16-bit to 8-bit graphics conversions), but with the noise already inherent in digital readouts before the ADC, it would seem unnecessary for 12-bit readout, to me.

if
an A-D converter has a noise floor much worse than the number of
bits it contains would suggest then it contains bits which are
essentially worthless (I would normally expect a real 12 bit
converter to have a DR near the range of an ideal 11 bit
converter). Although exaggerated bit depths are sometimes seen in
marketing blurbs from unscrupulous companies I don't expect that
the A-D converters in Canon's d-slrs suffer from it. I would
therefore expect that when they increase the bit depth by 2 bits
the dynamic range would increase by something near 2 stops. If the
analog input noise dominated the 12 bit converter noise and
remained the same in a 14 bit converter then Canon would be guilty
of wasteful (in terms of raw data storage) and misleading
specsmanship for the 1DIII.

Well, what did they claim? They claimed a higher bit depth, but did they claim greater DR at ISO 100?

There is still the possibility of higher DR at higher ISOs (higher than the mkII; not higher than mkIII ISO 100), even if IR's pre-production unit has the final readout quality at ISO 100.

14 bits will make the RAW files bigger, by a ratio of more than 14:12, as these extra bits are not as compressable, since they are more random than all the rest. It may, however, force converters to use more precision for WB, demosaicing, etc, giving subtle improvements in output (much like increasing the bit depth in photoshop before editing). If the banding noise is related somehow to bit depth, that may improve as well. The banding you see in deep shadows is often only + - 1 ADU, but is very visible nonetheless. After production units come out, it would be interesting to compare banding, as well as other noises, between the mkII and mkIII. Banding is one of the biggest problems with current Canon DSLRs, IMO, especially when you don't have enough light to expose to the right at ISO 1600.

--
John

 

[Its RKM]

Of course it is limited by analogue noise before quantisation -
nobody has suggested that the CMOS chip had infinite signal to
noise or dynamic range. However the data is further limited by the
ADC when ISO is low. Under those conditions it certainly IS
"bit-limited" as you put it. This is clearly demonstrated in Roger
Clarke's data and charts.

No, it is not.

Yes it is. Take a look at RC's table again. When the noise from the sensor equals the noise from the ADC the synamic range is more than 11 ENOBs. So for less than a stop lower than that the sustem certainly IS bit limited. Other noise sources might be significant too, but limited bits are certainyl limiting the dynamic range at lower ISOs.

What I have said
is that it is not the BIT DEPTH , which is a different thing.

No, that isn't what you said. Fortunately, precisely what you said and what you contested are recorded on the forum for eternity.

 

[John Sheehy]

I thought that more bits could mean smoother transitions, even over
the same DR, at least for JPEG?

It could, but if analog noise dominates, it can't work its wonders very well.

--
John

 

[Julia Borg]

you have some amount of light that hits a cell. that light is converted into electrons. the cell can hold only so much electrons, let's say 80.000.

now, you can convert 80.000 electrons through 14-bit ADC, or 8-bit ADC. it does not change dynamic range, dynamic range is determined by the maximum amount of electrones the cell can hold, and by noise.

you can measure the number in dozens, or in grosses. it does not change the number, it only changes granularity.

(when you are above at base ISO, depending on design, things can look differently and more bits in ADC can result in less highlight clipping.)

--
Julia

 

[ohyva]

"DR and noise go hand in hand" is true when we think how far the DR
extend to dark shades. But its possible also to improve the dynamic
range in how far you get usable shades in high-lights - today this
is about 3.5-4 stops over the mid-gray.

That's only a semantic, relativistic issue, though. The bottom
line is that the RAW data at any given ISO clips at a certain
level, and a signal with what you consider the "lowest usable S/N
ratio" lies a certain number of stops below that point; that is all
that determines dynamic range. It is irrelevant and arbitrary
where the middle-grey point is; it is only a frame of reference for
metering.

I do not agree, as IMHO the biggest problems today in dSLRs are the burned high-lights. The difference comparede to film is so visible.

dSLRs are tuned as a compromize, and I think the high-light preserving mode in 1D3 is a very good idea.

The same sensor with the same amplification could have an ISO 100
with 4 stops of highlights above middle grey, or it can be
considered ISO 200 with 5 stops, or ISO 50 with 3 stops.

Of couse yes, but I still think the sensor manufacturers should implement some high-light improvement. Pushing the noise lower and lower has limits and I do not know how much extra DR we can expect from there.

 

[Peter Carmichael]

"DR and noise go hand in hand" is true when we think how far the DR
extend to dark shades. But its possible also to improve the dynamic
range in how far you get usable shades in high-lights - today this
is about 3.5-4 stops over the mid-gray.

That's only a semantic, relativistic issue, though. The bottom
line is that the RAW data at any given ISO clips at a certain
level, and a signal with what you consider the "lowest usable S/N
ratio" lies a certain number of stops below that point; that is all
that determines dynamic range. It is irrelevant and arbitrary
where the middle-grey point is; it is only a frame of reference for
metering.

I do not agree, as IMHO the biggest problems today in dSLRs are the
burned high-lights. The difference comparede to film is so visible.

Whether you agree or not, John is absolutely right.

At the raw capture, the technology is agnostic of highlights, shadows and mid-range tones. It is a linear capture. Dynamic range is dynamic range. There are no two ways about it. Lowering the noise floor so that shadow information is retrievable is exactly the same as having highlight range. Anybody who tells you differently is in marketing.

In converting this information to an output file in a generic colorspace, the numbers then gain exact rendering intents. 117,117,117 in sRGB is 18% gray and the colourspace only has 2.5 stops of highlight range.

DSLRs are tuned as a compromize, and I think the high-light
preserving mode in 1D3 is a very good idea.

It simply exposes a bit further to the left.

The same sensor with the same amplification could have an ISO 100
with 4 stops of highlights above middle grey, or it can be
considered ISO 200 with 5 stops, or ISO 50 with 3 stops.

Of couse yes, but I still think the sensor manufacturers should
implement some high-light improvement. Pushing the noise lower and
lower has limits and I do not know how much extra DR we can expect
from there.

The "sensor" already has greater capability than we see in individual raw files. This thread has spent a lot of time wondering whether 14 bit raw files allow better dynamic range than 12 bit. John has been entirely correct in pointing out observations about the noise floor in Canon implementations and in particular the pattern read noise that might prevent the full potential of having 2 more bits being achieved. I have been more or less correct in pointing out that in a non-implementation specific sense the sensor data exceeds 12 bits and 14 bit raw, appropriately implemented, would be gigantic step forward.

The words "appropriately implemented" are the key here and they tacitly acknowledge John's standpoint.

 

[Peter Carmichael]

you have some amount of light that hits a cell. that light is
converted into electrons. the cell can hold only so much electrons,
let's say 80.000.

Yes.

now, you can convert 80.000 electrons through 14-bit ADC, or 8-bit
ADC. it does not change dynamic range

No. And the reason the answer is no is contained in the last three words of the remaining part of your sentence.

dynamic range is determined
by the maximum amount of electrones the cell can hold
, and by noise.

There we have it ladies and gentlemen. Julia Borg agrees that noise defines the lower bound of dynamic range.

We had better ask Julia what makes up noise levels in 8 bit and 14 bit linear digitizations and see what conclusions we can draw.