# sources on 13 EV DR possible from 12 bit sensor

Started Apr 19, 2012 | Discussions
 Forum
Re: Some more detail

boggis the cat wrote:

Yes, what you are missing is that it is this type of statistical process that is involved in attributing DxO-type DR numbers to sensors. Then they "normalise" to 8 MPixel which further bumps up the "DR". So you end up with the potential for a claimed DR higher than the raw bit-depth -- clearly incorrect at the per-pixel level.

Then use the "screen" measure for DR that DxOMark gives, which is the per-pixel measure.

My view is that the per-pixel DR capability is more important than having enough pixels to yield a big number via a DxO-type process. Ultimately it's a complex aspect to evaluate and is really a judgement call, and so we have divergent DR claims for the same sensor depending on the assumptions and methods used.

DR / pixel is irrelevant when comparing systems with different pixel counts -- DR / area is the relevant measure. Just as lp/mm is irrelevent in a resolution test -- lw/ph is the relevant measure.

As for the measure of DR, it's simple: the DR is the number of levels of light where detail can be recorded, and is calculated as the number of stops from the noise floor to the saturation limit.

If the bit depth is less than the DR, then not all levels of the DR will be distinctly represented. For example, a jpg can produce well over 8 stops of DR, but it can't represent more than 8 distinct levels.

Complain
Re: Fuji S5 Pro

Leo wrote:

Aleo Veuliah wrote:

I remember to read that the Fuji S5 Pro, has 14 EV DR, and I think it is a 12 bit

Fuji S5 (2006) and S3 (2004) DR was around 13.5EV. It is good now but that time the 13.5EV was a fantastic number!
Leo

You are right, just checked and it has 13.5, my memory told me 14, but I was wrong, anyway is still (and still is) a fantastic like you said

-- hide signature --

Simplicity is the ultimate sophistication.

God is the tangential point between zero and infinity.

Imagination is more important than knowledge.

God always take the simplest way.

Aleo Veuliah's gear list:Aleo Veuliah's gear list
Panasonic Lumix DMC-ZS1 Panasonic FZ1000 Panasonic Lumix DMC-GF1 Panasonic Lumix DMC-G2 Panasonic Lumix DMC-GH4 +5 more
Complain
Re: Lack of detail

Detail Man wrote:

boggis the cat wrote:

Detail Man wrote:

Aleo Veuliah wrote:

I looks complicated at the first view, thanks for the information

"Cutting to the chase" it represents an entirely bogus premise evidently born out of wistful fantasy

You appear to have inadvertently left out your explanation of why you believe the example to be incorrect.

Or, "cutting to the chase", you have omitted the detail.

Good to see that you are still smiling ! I tell you what. Build it and I may buy one the next time that I need to photograph patterns of ever-sparser point-sources that have somehow made it through all of the preceding optics without without diffraction, dispersions or other aberrations ...

I am also ignoring the fact that the net photo-site output signals involved would seem to be squarely placed beneath the smallest quantization-error level of a normal (where the output-data bits have binary weightings) linear analog/digital converter. That just could give one some problems

Or am I missing something important in the "details" involved? Do feel free to further elaborate!

Or more fundamentally:

boggis the cat wrote:

• Assume a "perfect" sensor (no noise) with 1000 pixels. Take a photo of a bright white source, such that the raw file contains values of "4095" for each pixel -- all 12 bits are set.

• Reduce the lighting to obtain a raw file with values of "2047" -- now 11 bits are set, so that's a -1 EV DR drop.

Of course, this is physically impossible. He hasn't heard of shot noise.

-- hide signature --

Bob

Complain
Lack of detail is no barrier to wistfully dreaming ... (This I know)

bobn2 wrote:

Detail Man wrote:

boggis the cat wrote:

Detail Man wrote:

Aleo Veuliah wrote:

I looks complicated at the first view, thanks for the information

"Cutting to the chase" it represents an entirely bogus premise evidently born out of wistful fantasy

You appear to have inadvertently left out your explanation of why you believe the example to be incorrect.

Or, "cutting to the chase", you have omitted the detail.

Good to see that you are still smiling ! I tell you what. Build it and I may buy one the next time that I need to photograph patterns of ever-sparser point-sources that have somehow made it through all of the preceding optics without diffraction, dispersions or other aberrations ...

I am also ignoring the fact that the net photo-site output signals involved would seem to be squarely placed beneath the smallest quantization-error level of a normal (where the output-data bits have binary weightings) linear analog/digital converter. That just could give one some problems

Or am I missing something important in the "details" involved? Do feel free to further elaborate!

Or more fundamentally:

boggis the cat wrote:

• Assume a "perfect" sensor (no noise) with 1000 pixels. Take a photo of a bright white source, such that the raw file contains values of "4095" for each pixel -- all 12 bits are set.

• Reduce the lighting to obtain a raw file with values of "2047" -- now 11 bits are set, so that's a -1 EV DR drop.

Of course, this is physically impossible. He hasn't heard of shot noise.

Actually, my friend, the protagonist has created a "special condition" invincible to all our criticisms:

Don't confuse a simple mathematical model with the complexities of reality.

I'm so embarrased/ashamed of myself now that I shall pack it in and go off and play with myself ...

I could build a mountain from all of the wadded-up pieces of engineering-paper that I have amassed over the years due to my stubborn insistence that conceptual elegance came before realizability - and one single math error can (and has more than once) led to the prettiest of fallacious dreams. That's why I am a lousy engineer - I am a mystic at heart in love with novelty, and only time, tides, and ridicule have taught me to always doubt myself, then doubt some more .

Only in menâ€™s imagination does every truth find an effective and undeniable existence .
Imagination, not invention, is the supreme master of art as of life .

Nevertheless, existence remains the great humbler,

DM ...

Complain
Re: Lack of detail

bobn2 wrote:
...........

boggis the cat wrote:

• Assume a "perfect" sensor (no noise) with 1000 pixels. Take a photo of a bright white source, such that the raw file contains values of "4095" for each pixel -- all 12 bits are set.

• Reduce the lighting to obtain a raw file with values of "2047" -- now 11 bits are set, so that's a -1 EV DR drop.

Of course, this is physically impossible. He hasn't heard of shot noise.

-- hide signature --

Bob

boggis the cat is correct. The half takes 2047 levels. Then divide the remaining 2047 by half and etc ... at the end you will find that there are not many tonal levels left for low light area, which directly effects details in low light.
Leo

Leo's gear list:Leo's gear list
Pentax K-S2 Pentax smc DA 16-45mm F4 ED AL HD Pentax DA 55-300mm F4.0-5.8 ED WR
Complain
Re: Lack of detail

Leo wrote:
bobn2 wrote:
...........

boggis the cat wrote:

• Assume a "perfect" sensor (no noise) with 1000 pixels. Take a photo of a bright white source, such that the raw file contains values of "4095" for each pixel -- all 12 bits are set.

• Reduce the lighting to obtain a raw file with values of "2047" -- now 11 bits are set, so that's a -1 EV DR drop.

Of course, this is physically impossible. He hasn't heard of shot noise.

-- hide signature --

Bob

boggis the cat is correct.

No, he isn't.

The half takes 2047 levels.

There is no way of making every pixel count to 2047. There is a tiny probability that it could just happen that way, but in practice there is a statistical variation from pixel to pixel caused by the quantum nature of the light. That is just physics. The standard deviation of that variation goes as the square root of the number of photons captured per pixel. To know what it is, you have to know how many photons '2047' represents.

Then divide the remaining 2047 by half and etc ... at the end you will find that there are not many tonal levels left for low light area, which directly effects details in low light.

There are not rigidly defined 'tonal levels' due to the noise, which is inherent in the structure of the light. Even at the lowest level, if a count of 1 represents 1 photon, then there is a variation of sqrt(1), which is 1. Even right at the lowest levels there is noise dithering going on. In practice, 2047 linear levels is overencoded, which is why Nikon lossy NEF and Sony raws use non-linear encoding, to eliminate the unneeded 'levels'.

-- hide signature --

Bob

Complain
Re: Lack of detail

Is not that what we called noise. For higher noise the DR is lower. For noise 1EV difference DR would be much wider then for 0.1EV. That what Imaging Resources always shows in their DR tests. Please correct me if wrong.
Leo

Leo's gear list:Leo's gear list
Pentax K-S2 Pentax smc DA 16-45mm F4 ED AL HD Pentax DA 55-300mm F4.0-5.8 ED WR
Complain
Re: An example that may make sense

boggis the cat wrote:

highwave wrote:
A 12-bit raw can store 2^12 or 4096 values -- 0 to 4095.

• Assume a "perfect" sensor (no noise) with 1000 pixels. Take a photo of a bright white source, such that the raw file contains values of "4095" for each pixel -- all 12 bits are set.

• Reduce the lighting to obtain a raw file with values of "2047" -- now 11 bits are set, so that's a -1 EV DR drop.

Only if luminance range (in the channel) encoded as 2047 is 1EV darker than luminance range encoded as 4095. Which might not be the case at all. And nobody besides the Olympus embedded software engineer who programmed the part really knows is it so or not.

Complain
Re: Sounds reasonable

Raist3d wrote:

if they did non linear mapping of bits I guess you could represent more too but tonality suffers. And it wouldn't be a trivial thing to do in a visually pleasing way.

It is trivial thing to do in a visually pleasing way, as people only able to recognize approx 450 different levels of gray in ideal laboratory condition on a big screen (9 bits are obviously more than enough to encode them all). Much less on a pixel level, much much less in individual color channels.

12 bits are only needed for postprocessing.

Complain
Re: It isn't a per-pixel measurement

cameron2 wrote:

boggis the cat wrote:

It's a statistical measurement based on an area of pixels. A 12-bit number obviously can't hold more than "12 EV" of data, so per-pixel DR is lower.

Technically, one could handle "13 EV of data" with only 4 bits (2^4=16).

Technically, absolutely any DR can be handled by just 2 bits. Encode up to 1 cd/m2 as '00', and 65536+ cd/m2 as '11', and or goodness, we have DR of 16 EV! Call the press!

With 12 bits of information, one can handle over 4000 levels of light, and that can be mapped (e.g. by a curve) to any values, positive or negative, of any range.

A lot of people refuse to understand that, no matter how many times they are told so.

Complain
It isn't a matter of armchair pontifications ...

peevee1 wrote:

Technically, absolutely any DR can be handled by just 2 bits.

Try 1 bit (pulse-width-modulation ADCs). The term "theoretically" would be a lot more apropos, as anybody who has actually ever attempted to build a real-life system soon becomes aware of the fact that "Mother Nature does not give willingly", and all machines are subject to pesky little non-ideal effects and limitations which summarily "kick one's ass". It's often called "engineering".

Complain
Re: sources on 13 EV DR possible from 12 bit sensor - Some Facts

I'd guess the Canon people are somewhat disappointed with the new sensor.

It only has to be better than 5D Mark II and D800. It is. Otherwise, they got 1D to sell.

Not a whole lot better than the 5D II in that department.

Whole stop better where it matters most.

Complain
Re: Lack of detail

Leo wrote:

Is not that what we called noise.

Really? What do you call noise, then?

For higher noise the DR is lower.

For a higher noise floor, the DR is lower. the noise floor is essentially the read noise, not the shot noise.

For noise 1EV difference DR would be much wider then for 0.1EV.

! stop read noise difference causes one stop DR difference.

That what Imaging Resources always shows in their DR tests.

What is what IR shows in their DR tests?

I'm don't know what point you are making, and how what you are saying related to whether as many bits as stops are needed to encode DR.

-- hide signature --

Bob

Complain
Re: It isn't a matter of armchair pontifications ...

Detail Man wrote:

peevee1 wrote:

Technically, absolutely any DR can be handled by just 2 bits.

Modulation has nothing to do with amount of information measured in bits.

Complain
Re: It isn't a matter of armchair pontifications ...

peevee1 wrote:

Detail Man wrote:

peevee1 wrote:

Technically, absolutely any DR can be handled by just 2 bits.

Modulation has nothing to do with amount of information measured in bits.

The issue here is modulation, not whether all the information is encoded.
--
Bob

Complain
Re: sources on 13 EV DR possible from 12 bit sensor - Some Facts

peevee1 wrote:

I'd guess the Canon people are somewhat disappointed with the new sensor.

It only has to be better than 5D Mark II and D800. It is.

It certainly isn't better than the D800. It is better than 5DII.
--
Bob

Complain
Re: sources on 13 EV DR possible from 12 bit sensor

highwave wrote:

While I'm very aware that limits and laws in physics are constantly being broken,

Complain
Re: It isn't a matter of armchair pontifications ...

bobn2 wrote:

peevee1 wrote:

Detail Man wrote:

peevee1 wrote:

Technically, absolutely any DR can be handled by just 2 bits.

Modulation has nothing to do with amount of information measured in bits.

The issue here is modulation, not whether all the information is encoded.

When discussing bits and raw files, it is the issue of information, not modulation (which only applies to analogue signal).

Complain
Re: sources on 13 EV DR possible from 12 bit sensor - Some Facts

bobn2 wrote:

peevee1 wrote:

I'd guess the Canon people are somewhat disappointed with the new sensor.

It only has to be better than 5D Mark II and D800. It is.

It certainly isn't better than the D800. It is better than 5DII.

http://www.sensorgen.info/CanonEOS_5D_MkIII.html
ISO Measured ISO Read Noise (e-) Saturation (e-) DR (stops)
12800 10084 2.4 526 7.8
25600 18589 2.9 308 6.7

http://www.sensorgen.info/NikonD800.html
ISO Measured ISO Read Noise (e-) Saturation (e-) DR (stops)
12800 8661 3.3 507 7.3
25600 16117 3.7 252 6.1

Complain
Re: sources on 13 EV DR possible from 12 bit sensor

Timur Born wrote:

What you meant is 72 dB S/N ratio, each bit representing 6(.02) dB, which in turn each represent a doubling of the signal. Each doubling of the signal means 1 EV. 72/6 = 12.

Doubling is ~3 dB.

-- hide signature --
Complain
 Forum