Engineering dynamic range vs dynamic range

Sure, but what number (of what measure) told you it was 'enough'?

My personal opinion is that once you apply tone curves to images, any DR much over 14 EV is pointless. You are more likely to be restricted by mid-tone SNR than DR,
You can plug any definition that you want for acceptable SNR into one of my shadow noise curves, and you'll get the DR associated with that threshold.

https://blog.kasson.com/the-last-word/fundamental-camera-noise-parameters/
A very useful analysis, thankyou.

I would be interested to see the same curves adjusted to match a standard ACR tone-curve for those cameras. This tends to amplify low-mid to mid-tone noise and supress shadow noise, hence my impression that it's often a limiting factor when you reduce exposure - as long as read noise is still buried in the toe of the curve.
With the dynamic range of today's cameras, you're not going to see shadow noise at low ISOs unless you significantly boost the shadows, so the default tone curve is not relevant here.
I would think such an analysis for various degrees of underexposure might reveal a maximum in the curve somewhere between 0 and 128, which would represent the useful exposure latitude limit.
I'm not clear on the curve you're referring to.
With any reasonable SNR threshold for photographically-interesting non-documentary work, I've never seen a camera have close to a 14-stop DR.
That's entirely reasonable - I cited 14 EV because it applies to the EDR of several cameras as measured by DXO - which are normalised to 8 MP. The useful value may be lower.

I have downloaded images from the cameras concerned and in all cases there are some other factors affecting the appearance of noise near the black level. Flare, raw noise smoothing, blotching, colour casts, banding, etc. The appearance is not accurately predicted by the noise, per se. Hence my contention that of all the sensor metrics, this is the most misleading and the hardest to pin down in a real-world context.
It depends on the threshold. There is a 1/f component to most cameras' RN, and it tends to be static. But photon noise is white, and when you get to the point where photon noise dominates read noise (remember, they add as square root of the sum of the squares), and before you get to the point where photon noise is overwhelmed by PRNU, then we're dealing with white noise.
 
57even wrote:
...
No useful information can be defined as the point where SNR=1 (signal=noise)
Not really directed at you but every mention of SNR=1 in this thread is incorrect.

It's careless shorthand for Signal = Read Noise
Not when I say SNR = 1. All noise is not read noise.
Just curious, under what circumstance is SNR=-1 (as opposed to signal = read noise) relevant? Surely you're not saying this is the lower bound for EDR (as opposed to read noise)?
I am not saying that. Some people use SNR = 1 for EDR. I am not one of those people; I find it difficult to measure compared to RN; it needs a PTC to get the measurement. I do PTCs, but I don't use then for EDR.
Is it not possible to calculate it, if we assume short exposures (no dark signal) and we know the read noise?
 
57even wrote:
...
No useful information can be defined as the point where SNR=1 (signal=noise)
Not really directed at you but every mention of SNR=1 in this thread is incorrect.

It's careless shorthand for Signal = Read Noise
Not when I say SNR = 1. All noise is not read noise.
Just curious, under what circumstance is SNR=-1 (as opposed to signal = read noise) relevant? Surely you're not saying this is the lower bound for EDR (as opposed to read noise)?
I am not saying that. Some people use SNR = 1 for EDR. I am not one of those people; I find it difficult to measure compared to RN; it needs a PTC to get the measurement. I do PTCs, but I don't use then for EDR.
Is it not possible to calculate it, if we assume short exposures (no dark signal) and we know the read noise?
Not sure what you mean by dark signal there. Ignoring that, sure. We have to know the FWC as well, of course. But how do we get the FWC? By doing a PTC. Catch 22.
 
Sure, but what number (of what measure) told you it was 'enough'?

My personal opinion is that once you apply tone curves to images, any DR much over 14 EV is pointless. You are more likely to be restricted by mid-tone SNR than DR,
You can plug any definition that you want for acceptable SNR into one of my shadow noise curves, and you'll get the DR associated with that threshold.

https://blog.kasson.com/the-last-word/fundamental-camera-noise-parameters/
A very useful analysis, thankyou.

I would be interested to see the same curves adjusted to match a standard ACR tone-curve for those cameras. This tends to amplify low-mid to mid-tone noise and supress shadow noise, hence my impression that it's often a limiting factor when you reduce exposure - as long as read noise is still buried in the toe of the curve.
With the dynamic range of today's cameras, you're not going to see shadow noise at low ISOs unless you significantly boost the shadows, so the default tone curve is not relevant here.
I agree, but exposure latitude is not only dependent on shadow noise. You are losing 1 stop of SNR for each -1EV of exposure compensation.
I would think such an analysis for various degrees of underexposure might reveal a maximum in the curve somewhere between 0 and 128, which would represent the useful exposure latitude limit.
I'm not clear on the curve you're referring to.
Your curve adjusted for output contrast. You could normalise it to slope=1 at mid-tone, since we are only looking for a peak value.
With any reasonable SNR threshold for photographically-interesting non-documentary work, I've never seen a camera have close to a 14-stop DR.
That's entirely reasonable - I cited 14 EV because it applies to the EDR of several cameras as measured by DXO - which are normalised to 8 MP. The useful value may be lower.

I have downloaded images from the cameras concerned and in all cases there are some other factors affecting the appearance of noise near the black level. Flare, raw noise smoothing, blotching, colour casts, banding, etc. The appearance is not accurately predicted by the noise, per se. Hence my contention that of all the sensor metrics, this is the most misleading and the hardest to pin down in a real-world context.
It depends on the threshold. There is a 1/f component to most cameras' RN, and it tends to be static. But photon noise is white, and when you get to the point where photon noise dominates read noise (remember, they add as square root of the sum of the squares), and before you get to the point where photon noise is overwhelmed by PRNU, then we're dealing with white noise.
Sure, I get you. All I am saying is that in real images, there are other factors at play that mess with our perception at low signal levels. Often, they are more distracting than the actual noise, so DR is not very relevant.
 
57even wrote:
...
No useful information can be defined as the point where SNR=1 (signal=noise)
Not really directed at you but every mention of SNR=1 in this thread is incorrect.
Not when I use it. Some shot noise is included since S component is non zero.
It's careless shorthand for Signal = Read Noise
SNR=1 is a different place as you can tell by inspecting any Photon Transfer Curve (PTC).
as in the ISO specification
I'm unaware of any ISO specification that refers to SNR=1
ISO 15739

But also EMVA 1288.

Note, the value SNR=1 is read directly off the response curve.
Interesting, thanks. But not related to EDR; right.
It's one way to measure it.
I'm not a believer in their being more than one way.
I think DXO also use SNR=1.
FWIW, DxO uses read noise not SNR=1.
Imatest has a whole slew of DR measurement methods - ISO, EMVA, contrast, etc.

Is any of them an agreed definition of EDR? AFAIK, there isn't one.

Is this part of the problem perhaps?
I guess so.
 
57even wrote:
...
No useful information can be defined as the point where SNR=1 (signal=noise)
Not really directed at you but every mention of SNR=1 in this thread is incorrect.

It's careless shorthand for Signal = Read Noise
Not when I say SNR = 1. All noise is not read noise.
Just curious, under what circumstance is SNR=-1 (as opposed to signal = read noise) relevant? Surely you're not saying this is the lower bound for EDR (as opposed to read noise)?
I am not saying that. Some people use SNR = 1 for EDR. I am not one of those people; I find it difficult to measure compared to RN; it needs a PTC to get the measurement. I do PTCs, but I don't use then for EDR.
Is it not possible to calculate it, if we assume short exposures (no dark signal) and we know the read noise?
Not sure what you mean by dark signal there. Ignoring that, sure. We have to know the FWC as well, of course. But how do we get the FWC? By doing a PTC. Catch 22.
Right, although you can get FWC (gain) without doing a PTC.
 
57even wrote:
...
No useful information can be defined as the point where SNR=1 (signal=noise)
Not really directed at you but every mention of SNR=1 in this thread is incorrect.

It's careless shorthand for Signal = Read Noise
Not when I say SNR = 1. All noise is not read noise.
Just curious, under what circumstance is SNR=-1 (as opposed to signal = read noise) relevant? Surely you're not saying this is the lower bound for EDR (as opposed to read noise)?
I am not saying that. Some people use SNR = 1 for EDR. I am not one of those people; I find it difficult to measure compared to RN; it needs a PTC to get the measurement. I do PTCs, but I don't use then for EDR.
Is it not possible to calculate it, if we assume short exposures (no dark signal) and we know the read noise?
Not sure what you mean by dark signal there. Ignoring that, sure. We have to know the FWC as well, of course. But how do we get the FWC? By doing a PTC. Catch 22.
Ah, sorry = I was only referring to the point where SNR = 1.

If you assume the noise is only shot plus read noise...

S = N implies that S^2 = S+N^2, and since N^2 is a constant, it's a simple quadratic. Just solve for S and insert read noise back in.

Yes, to calculate DR you need FWC as well.
 
57even wrote:
...
No useful information can be defined as the point where SNR=1 (signal=noise)
Not really directed at you but every mention of SNR=1 in this thread is incorrect.
Not when I use it. Some shot noise is included since S component is non zero.
It's careless shorthand for Signal = Read Noise
SNR=1 is a different place as you can tell by inspecting any Photon Transfer Curve (PTC).
as in the ISO specification
I'm unaware of any ISO specification that refers to SNR=1
ISO 15739

But also EMVA 1288.

Note, the value SNR=1 is read directly off the response curve.
Interesting, thanks. But not related to EDR; right.
It's one way to measure it.
I'm not a believer in their being more than one way.
I think DXO also use SNR=1.
FWIW, DxO uses read noise not SNR=1.
Good to know that, thanks! I wish they hadn't removed their methodology pages!!!!
Imatest has a whole slew of DR measurement methods - ISO, EMVA, contrast, etc.

Is any of them an agreed definition of EDR? AFAIK, there isn't one.

Is this part of the problem perhaps?
I guess so.
As I said, there are only 2 standards I know of - ISO and EMVA. Both are based on a PTC.

The issue with ISO is that is has to be corrected for density.

Lot's of good stuff on the Imatest Website. Interesting reading and perhaps thought provoking?
 
Sure, but what number (of what measure) told you it was 'enough'?

My personal opinion is that once you apply tone curves to images, any DR much over 14 EV is pointless. You are more likely to be restricted by mid-tone SNR than DR,
You can plug any definition that you want for acceptable SNR into one of my shadow noise curves, and you'll get the DR associated with that threshold.

https://blog.kasson.com/the-last-word/fundamental-camera-noise-parameters/
A very useful analysis, thankyou.

I would be interested to see the same curves adjusted to match a standard ACR tone-curve for those cameras. This tends to amplify low-mid to mid-tone noise and supress shadow noise, hence my impression that it's often a limiting factor when you reduce exposure - as long as read noise is still buried in the toe of the curve.
With the dynamic range of today's cameras, you're not going to see shadow noise at low ISOs unless you significantly boost the shadows, so the default tone curve is not relevant here.
I agree, but exposure latitude is not only dependent on shadow noise. You are losing 1 stop of SNR for each -1EV of exposure compensation.
I would think such an analysis for various degrees of underexposure might reveal a maximum in the curve somewhere between 0 and 128, which would represent the useful exposure latitude limit.
I'm not clear on the curve you're referring to.
Your curve adjusted for output contrast. You could normalise it to slope=1 at mid-tone, since we are only looking for a peak value.
Give me the actual math you wish me to perform. I'm still not getting it.
With any reasonable SNR threshold for photographically-interesting non-documentary work, I've never seen a camera have close to a 14-stop DR.
That's entirely reasonable - I cited 14 EV because it applies to the EDR of several cameras as measured by DXO - which are normalised to 8 MP. The useful value may be lower.

I have downloaded images from the cameras concerned and in all cases there are some other factors affecting the appearance of noise near the black level. Flare, raw noise smoothing, blotching, colour casts, banding, etc. The appearance is not accurately predicted by the noise, per se. Hence my contention that of all the sensor metrics, this is the most misleading and the hardest to pin down in a real-world context.
It depends on the threshold. There is a 1/f component to most cameras' RN, and it tends to be static. But photon noise is white, and when you get to the point where photon noise dominates read noise (remember, they add as square root of the sum of the squares), and before you get to the point where photon noise is overwhelmed by PRNU, then we're dealing with white noise.
Sure, I get you. All I am saying is that in real images, there are other factors at play that mess with our perception at low signal levels. Often, they are more distracting than the actual noise, so DR is not very relevant.
I guess we'll have to disagree on the DR is not very relevant part.
 
57even wrote:
...
No useful information can be defined as the point where SNR=1 (signal=noise)
Not really directed at you but every mention of SNR=1 in this thread is incorrect.

It's careless shorthand for Signal = Read Noise
Not when I say SNR = 1. All noise is not read noise.
Just curious, under what circumstance is SNR=-1 (as opposed to signal = read noise) relevant? Surely you're not saying this is the lower bound for EDR (as opposed to read noise)?
I am not saying that. Some people use SNR = 1 for EDR. I am not one of those people; I find it difficult to measure compared to RN; it needs a PTC to get the measurement. I do PTCs, but I don't use then for EDR.
Is it not possible to calculate it, if we assume short exposures (no dark signal) and we know the read noise?
Not sure what you mean by dark signal there. Ignoring that, sure. We have to know the FWC as well, of course. But how do we get the FWC? By doing a PTC. Catch 22.
Right, although you can get FWC (gain) without doing a PTC.
You're right, you don't have to do the whole PTC; you can spot check. But I usually go ahead and do the PTC to get the FWC.
 
FWIW, DxO uses read noise not SNR=1.
Bill, many moons ago there was a page at DXOmark.com that explained that they calculate 'Screen' DR from 'full SNR curves', from clipping to SNR=1, so from full scale to 0dB on the x axis below (these are also the curves that they use to extract the SNR18% figures):

Nikon Z7II Full SNR Curves from Measurement tab at dxomark.com
Nikon Z7II Full SNR Curves from Measurement tab at dxomark.com

Alas I can't find it in their reworked site. Note however that the base ISO PTC for the Z7II hits SNR = 1 (0dB) at a rounded 0.009% of full scale (100%), equal to a DR of 13.44. The 'screen' tab on their Z7II DR page shows 13.48, the small difference due to the rounding error in the displayed number.

I believe these curves represent the average performance of the 4 color channels since they say that they come from mean gray levels, but I was never able to confirm this.

Jack

PS As I think we all know here, 'Print' DR is 'Screen' DR normalized to an 8MP image.
 
Last edited:
FWIW, DxO uses read noise not SNR=1.
Bill, many moons ago there was a page at DXOmark.com that explained that they calculate 'Screen' DR from 'full SNR curves', from clipping to SNR=1, so from full scale to 0dB on the x axis below (these are also the curves that they use to extract the SNR18% figures):

Nikon Z7II Full SNR Curves from Measurement tab at dxomark.com
Nikon Z7II Full SNR Curves from Measurement tab at dxomark.com

Alas I can't find it in their reworked site. Note however that the base ISO PTC for the Z7II hits SNR = 1 (0dB) at a rounded 0.009% of full scale (100%), equal to a DR of 13.44. The 'screen' tab on their Z7II DR page shows 13.48, the small difference due to the rounding error in the displayed number.

I believe these curves represent the average performance of the 4 color channels since they say that they come from mean gray levels, but I was never able to confirm this.

Jack

PS As I think we all know here, 'Print' DR is 'Screen' DR normalized to an 8MP image.
My recollection is that when I used to compare DxOMark to PhotonsToPhotos that it appeared to be read noise rather than SNR=1.

But perhaps my recollection is wrong!

FWIW eye-balling my Photon Transfer Curve (PTC) that 13.4x figure looks like SNR=1 (!).

27e057172fa5447a90829c7a1753fcdd.jpg.png

Yeah, I found it, they do say "a signal-to-noise ratio below 0 dB"

--
Bill ( Your trusted source for independent sensor data at PhotonsToPhotos )
 
Last edited:
Sure, but what number (of what measure) told you it was 'enough'?

My personal opinion is that once you apply tone curves to images, any DR much over 14 EV is pointless. You are more likely to be restricted by mid-tone SNR than DR,
You can plug any definition that you want for acceptable SNR into one of my shadow noise curves, and you'll get the DR associated with that threshold.

https://blog.kasson.com/the-last-word/fundamental-camera-noise-parameters/
A very useful analysis, thankyou.

I would be interested to see the same curves adjusted to match a standard ACR tone-curve for those cameras. This tends to amplify low-mid to mid-tone noise and supress shadow noise, hence my impression that it's often a limiting factor when you reduce exposure - as long as read noise is still buried in the toe of the curve.
With the dynamic range of today's cameras, you're not going to see shadow noise at low ISOs unless you significantly boost the shadows, so the default tone curve is not relevant here.
I agree, but exposure latitude is not only dependent on shadow noise. You are losing 1 stop of SNR for each -1EV of exposure compensation.
I would think such an analysis for various degrees of underexposure might reveal a maximum in the curve somewhere between 0 and 128, which would represent the useful exposure latitude limit.
I'm not clear on the curve you're referring to.
Your curve adjusted for output contrast. You could normalise it to slope=1 at mid-tone, since we are only looking for a peak value.
Give me the actual math you wish me to perform. I'm still not getting it.
NSR (reciprocal SNR) is affected directly by contrast. If you double the slope of a tone curve at one point, you effectively double the apparent NSR of in that region of the image. Or, conversely, halve the SNR.

This does changes the areas in the image where we tend to see the most noise.

From observations of grey step wedge in the DPR test charts, I reckoned the visible noise peaked at around RGB 40-60.

So, a plot of NSR/RGB corrected for contrast would show a peak at that level, or conversely SNR would have a minimum. But it doesn't have to be RGB, or NSR.

In theory, SNR/Signal should show a minimum at the same level.

With any reasonable SNR threshold for photographically-interesting non-documentary work, I've never seen a camera have close to a 14-stop DR.
That's entirely reasonable - I cited 14 EV because it applies to the EDR of several cameras as measured by DXO - which are normalised to 8 MP. The useful value may be lower.

I have downloaded images from the cameras concerned and in all cases there are some other factors affecting the appearance of noise near the black level. Flare, raw noise smoothing, blotching, colour casts, banding, etc. The appearance is not accurately predicted by the noise, per se. Hence my contention that of all the sensor metrics, this is the most misleading and the hardest to pin down in a real-world context.
It depends on the threshold. There is a 1/f component to most cameras' RN, and it tends to be static. But photon noise is white, and when you get to the point where photon noise dominates read noise (remember, they add as square root of the sum of the squares), and before you get to the point where photon noise is overwhelmed by PRNU, then we're dealing with white noise.
Sure, I get you. All I am saying is that in real images, there are other factors at play that mess with our perception at low signal levels. Often, they are more distracting than the actual noise, so DR is not very relevant.
I guess we'll have to disagree on the DR is not very relevant part.
Well, I guess it depends what criterion you use for minimum "usable" signal.
 
FWIW, DxO uses read noise not SNR=1.
Bill, many moons ago there was a page at DXOmark.com that explained that they calculate 'Screen' DR from 'full SNR curves', from clipping to SNR=1, so from full scale to 0dB on the x axis below (these are also the curves that they use to extract the SNR18% figures):

Nikon Z7II Full SNR Curves from Measurement tab at dxomark.com
Nikon Z7II Full SNR Curves from Measurement tab at dxomark.com

Alas I can't find it in their reworked site. Note however that the base ISO PTC for the Z7II hits SNR = 1 (0dB) at a rounded 0.009% of full scale (100%), equal to a DR of 13.44. The 'screen' tab on their Z7II DR page shows 13.48, the small difference due to the rounding error in the displayed number.

I believe these curves represent the average performance of the 4 color channels since they say that they come from mean gray levels, but I was never able to confirm this.

Jack

PS As I think we all know here, 'Print' DR is 'Screen' DR normalized to an 8MP image.
My recollection is that when I used to compare DxOMark to PhotonsToPhotos that it appeared to be read noise rather than SNR=1.

But perhaps my recollection is wrong!

FWIW eye-balling my Photon Transfer Curve (PTC) that 13.4x figure looks like SNR=1 (!).

27e057172fa5447a90829c7a1753fcdd.jpg.png

Yeah, I found it, they do say "a signal-to-noise ratio below 0 dB"
That was my recollection...

Glad we cleared that up ;-)

--
"A designer knows he has achieved perfection not when there is nothing left to add, but when there is nothing left to take away." Antoine de Saint-Exupery
 
Your curve adjusted for output contrast. You could normalise it to slope=1 at mid-tone, since we are only looking for a peak value.
Give me the actual math you wish me to perform. I'm still not getting it.
NSR (reciprocal SNR) is affected directly by contrast. If you double the slope of a tone curve at one point, you effectively double the apparent NSR of in that region of the image. Or, conversely, halve the SNR.

This does changes the areas in the image where we tend to see the most noise.

From observations of grey step wedge in the DPR test charts, I reckoned the visible noise peaked at around RGB 40-60.

So, a plot of NSR/RGB corrected for contrast would show a peak at that level, or conversely SNR would have a minimum. But it doesn't have to be RGB, or NSR.

In theory, SNR/Signal should show a minimum at the same level.
I don't see the actual math you wish me to perform in your response.
  1. What is the vertical axis?
  2. What is the horizontal axis?
  3. What is the metric for contrast?
 
That's the SNR after the image is scaled to a print size that is common for all cameras.

EDR is usually not normalized, so when comparting whatever threshold you're using for EDR to the threshold used for PDR, you need to consider the height of the sensor in pixels.
If you could make a sensor of arbitrary size using the same basic sensor elements, if you have a 100 pixel or a 100,000,000 pixel sensor isn't the SNR identical?

-- Bob
http://bob-o-rama.smugmug.com -- Photos
http://www.vimeo.com/boborama/videos -- Videos
 
That's the SNR after the image is scaled to a print size that is common for all cameras.

EDR is usually not normalized, so when comparting whatever threshold you're using for EDR to the threshold used for PDR, you need to consider the height of the sensor in pixels.
If you could make a sensor of arbitrary size using the same basic sensor elements, if you have a 100 pixel or a 100,000,000 pixel sensor isn't the SNR identical?
At the pixel level, yes. At the print level, no.
 
Your curve adjusted for output contrast. You could normalise it to slope=1 at mid-tone, since we are only looking for a peak value.
Give me the actual math you wish me to perform. I'm still not getting it.
NSR (reciprocal SNR) is affected directly by contrast. If you double the slope of a tone curve at one point, you effectively double the apparent NSR of in that region of the image. Or, conversely, halve the SNR.

This does changes the areas in the image where we tend to see the most noise.

From observations of grey step wedge in the DPR test charts, I reckoned the visible noise peaked at around RGB 40-60.

So, a plot of NSR/RGB corrected for contrast would show a peak at that level, or conversely SNR would have a minimum. But it doesn't have to be RGB, or NSR.

In theory, SNR/Signal should show a minimum at the same level.
I don't see the actual math you wish me to perform in your response.
  1. What is the vertical axis?
  2. What is the horizontal axis?
  3. What is the metric for contrast?
1. Visible NSR/SNR (NSR * tone curve gradient at relevant RGB).

2. RGB value - or saturation scaled to gamma 2.2

3. There isn't one. It's relative based on gradient.

--
"A designer knows he has achieved perfection not when there is nothing left to add, but when there is nothing left to take away." Antoine de Saint-Exupery
 
Last edited:
Your curve adjusted for output contrast. You could normalise it to slope=1 at mid-tone, since we are only looking for a peak value.
Give me the actual math you wish me to perform. I'm still not getting it.
NSR (reciprocal SNR) is affected directly by contrast. If you double the slope of a tone curve at one point, you effectively double the apparent NSR of in that region of the image. Or, conversely, halve the SNR.

This does changes the areas in the image where we tend to see the most noise.

From observations of grey step wedge in the DPR test charts, I reckoned the visible noise peaked at around RGB 40-60.

So, a plot of NSR/RGB corrected for contrast would show a peak at that level, or conversely SNR would have a minimum. But it doesn't have to be RGB, or NSR.

In theory, SNR/Signal should show a minimum at the same level.
I don't see the actual math you wish me to perform in your response.
  1. What is the vertical axis?
  2. What is the horizontal axis?
  3. What is the metric for contrast?
1. Visible NSR/SNR (NSR * tone curve gradient at relevant RGB).

2. RGB value - or saturation scaled to gamma 2.2

3. There isn't one. It's relative based on gradient.
How about the nonlinearities in the human visual system? Do they whole thing with the first derivative of L*?
 

Keyboard shortcuts

Back
Top