Connecting Exposure, ISO and Noise

[Panopeeper]

I am sure the vast majority of posters of these forums understand the connection between exposure, ISO and noise, but there may be many readers believing popular misconceptions like "my shots are noisy even at ISO 100", or "look how few noise is in this shot at ISO 12800".

I put a demonstration together showing the relationship between noise, exposure and ISO: The Source Of Noise

I ask for constructive critique of the content, form as well as language.


Thanks

[Luis Argerich]

Great article, very clear and I love the demonstrations. I know one shouldn't have the need to demonstrate the truth but people many times say "I don't believe it" so the examples are great.

thanks!
Luigi

[PeterAit]

Great article, very clear and I love the demonstrations. I know one shouldn't have the need to demonstrate the truth but people many times say "I don't believe it" so the examples are great.

thanks!
Luigi

And sometimes people don't believe the truth even after you demonstrate it!

Peter

[Guillermo Luijk]

Hi Gabor. I find the article very good if one already knows the interconnection between exposure, ISO and noise, but not so good for those who still don't understand it. You demonstrate the relations between those variables rather than explain them.

If I were to measure noise on my RAW files and would like to use RAWnalyze, this document would be just the perfect source of information.

However for those who are still finding out the connections you try to explain, the article could be confusing. The main problem I always found when talking to people about these subject is:

"Increasing the ISO setting decreases the noise up to a certain ISO level; further increasing the ISO setting is practically useless for the raw image."

This is a very advanced conclusion for someone who believed until now that: "the higher the ISO the more noise you get".

People are used to set ISO on their cameras and have apperture/shutter automatically adjusted by them. In that situation the variable ISO has not been isolated (which is actually the right way to analyse it), and I think this is the first thing that must be explained before saying to anyone that the higher the ISO, the higher the SNR. Otherwise they will not understand what you're talking about IMO.

BR

[dwdallam]

Great article, very clear and I love the demonstrations. I know one shouldn't have the need to demonstrate the truth but people many times say "I don't believe it" so the examples are great.

thanks!
Luigi

Well people contradict their own beliefs in spite of truth all of the time. For instance, in the USA most people believe in the Constitutional phrase that "All men are created equal," which really means that all people deserve the same treatment. The same people who believed the equality phrase took until 1921 to see that women must be allowed to vote, or they were holding contradictory positions. The "Truth" I speak of here is not in the phrase itself, but is in the contradiction. Even though all contradictions are false (which is true), people still hold on to contradictory belief systems.

[dwdallam]

Hi Gabor. I find the article very good if one already knows the interconnection between exposure, ISO and noise, but not so good for those who still don't understand it. You demonstrate the relations between those variables rather than explain them.

If I were to measure noise on my RAW files and would like to use RAWnalyze, this document would be just the perfect source of information.

However for those who are still finding out the connections you try to explain, the article could be confusing. The main problem I always found when talking to people about these subject is:

"Increasing the ISO setting decreases the noise up to a certain ISO level; further increasing the ISO setting is practically useless for the raw image."

This is a very advanced conclusion for someone who believed until now that: "the higher the ISO the more noise you get".

People are used to set ISO on their cameras and have apperture/shutter automatically adjusted by them. In that situation the variable ISO has not been isolated (which is actually the right way to analyse it), and I think this is the first thing that must be explained before saying to anyone that the higher the ISO, the higher the SNR. Otherwise they will not understand what you're talking about IMO.

BR

That sounds like a very astute and valid objection and observation.

[bjanes]

However for those who are still finding out the connections you try to explain, the article could be confusing. The main problem I always found when talking to people about these subject is:

"Increasing the ISO setting decreases the noise up to a certain ISO level; further increasing the ISO setting is practically useless for the raw image."

This is a very advanced conclusion for someone who believed until now that: "the higher the ISO the more noise you get".

People are used to set ISO on their cameras and have apperture/shutter automatically adjusted by them. In that situation the variable ISO has not been isolated (which is actually the right way to analyse it), and I think this is the first thing that must be explained before saying to anyone that the higher the ISO, the higher the SNR. Otherwise they will not understand what you're talking about IMO.

BR

This is an interesting dialog between two of the most knowledgeable and perceptive forum members. I agree with Guillermo that Gabor has brilliantly demonstrated some basic characteristics of noise and its sources, but without some theoretical considerations of the major sources of noise--read noise, shot noise, and PRNU (pixel response non uniformity)-- most readers will have trouble following the analysis. I would suggest that interested persons should read Emil Martinec's essay on noise and dynamic range in conjunction with Gabor's demonstration.

For a given exposure, the effect of ISO on noise is related to the contribution of read noise. Read noise is further complicated by whether it is expressed in electrons or ADUs, which are related to the sensor gain. Read noise can be reduced by increasing the ISO up to a certain point, beyond which ISO makes little difference. Gabor's demonstration is for a Canon camera, and the relationship between ISO and read noise is different for Canon and Nikon cameras as shown in Figures 15 and 15 of Emil's paper. It would be interesting to repeat the analysis for a Nikon D3 or other latest generation Nikon DSRL.

As Gabor has shown, noise in the color channels is related to exposure. Since current sensors are most sensitive to green light, the red and blue channels are relatively under-exposed with daylight exposures. With tungsten exposures the blue channel may be severely underexposed, leading to more noise. This may cause confusion for some.

Bill

[marcmccalmont]

I sure would like to know for a given situation which settings would give me minimum noise!
If I have a choice for a longer exposure but at a lower ISO where is my minimum noise?
When I lose an (or all) engines in an airplane, the flight management computer calculates an optimum airspeed to maximize my drift down distance.
Why can't my camera (or chart) calculate my optimum settings for minimum noise? I know ISO 100 for daytime shots, but in low light where am I better off? 1/2 sec @ISO 200 or 1/8 sec @ ISO 800? I would like to see some sort of "Minimum Noise Curve" say ISO vs. exposure time?
Marc

[Luis Argerich]

Longer exposure is always the way to go -If possible- in your example go with 1/2 sec if you can. The more light the sensor can get the better the signal/noise ratio will be.
The problem is that a long exposure is not always possible because subjects move, you may be without a tripod, etc. Only when you can't increase the exposure you can push the ISO to reduce noise.
Your camera has no idea if you are shooting a landscape or a person, handheld or tripod mounted so it can't take those decisions for you.

Luigi

I sure would like to know for a given situation which settings would give me minimum noise!
If I have a choice for a longer exposure but at a lower ISO where is my minimum noise?
When I lose an (or all) engines in an airplane, the flight management computer calculates an optimum airspeed to maximize my drift down distance.
Why can't my camera (or chart) calculate my optimum settings for minimum noise? I know ISO 100 for daytime shots, but in low light where am I better off? 1/2 sec @ISO 200 or 1/8 sec @ ISO 800? I would like to see some sort of "Minimum Noise Curve" say ISO vs. exposure time?
Marc

[ejmartin]

I sure would like to know for a given situation which settings would give me minimum noise!
If I have a choice for a longer exposure but at a lower ISO where is my minimum noise?
When I lose an (or all) engines in an airplane, the flight management computer calculates an optimum airspeed to maximize my drift down distance.
Why can't my camera (or chart) calculate my optimum settings for minimum noise? I know ISO 100 for daytime shots, but in low light where am I better off? 1/2 sec @ISO 200 or 1/8 sec @ ISO 800? I would like to see some sort of "Minimum Noise Curve" say ISO vs. exposure time?
Marc

The basic rule of thumb in most situations is

1) Choose the minimum aperture for the required depth of field.
2) Choose the minimum shutter speed required to freeze the action, prevent camera shake, etc.
3) Choose the ISO such that the histogram is pushed to the right as much as possible while keeping the highlights that you want to preserve unclipped.

There is one major exception to this rule of thumb, when item (3) suggests an ISO higher than about 1600.  In this case, one is usually better off underexposing at ISO 1600 at the aperture and shutter speed suggested by (1) and (2), since there is little to no improvement in S/N above ISO on most current cameras.  The threshold for making such a decision varies from camera model to camera model; for the 5D2, which has a lot of high ISO banding, it may make sense to continue raising the ISO to 3200 in step (3), since it lowers the banding noise relative to signal, while for the new Sony sensors such as in the D300 and D3x, the benefits of raising the ISO tail off somewhere between ISO 800 and 1600.

There are also a few footnotes to the above, such as the fact that there is almost nothing to be gained from using "intermediate" ISO's (the 1/3 stops in between ISO 100/200/400/800/1600 etc).

[Luis Argerich]

#1 should be maximum aperture (lowest F-number)
#2 should be maximum shutter speed (the longer the better)

The basic rule of thumb in most situations is

1) Choose the minimum aperture for the required depth of field.
2) Choose the minimum shutter speed required to freeze the action, prevent camera shake, etc.
3) Choose the ISO such that the histogram is pushed to the right as much as possible while keeping the highlights that you want to preserve unclipped.

There is one major exception to this rule of thumb, when item (3) suggests an ISO higher than about 1600.  In this case, one is usually better off underexposing at ISO 1600 at the aperture and shutter speed suggested by (1) and (2), since there is little to no improvement in S/N above ISO on most current cameras.  The threshold for making such a decision varies from camera model to camera model; for the 5D2, which has a lot of high ISO banding, it may make sense to continue raising the ISO to 3200 in step (3), since it lowers the banding noise relative to signal, while for the new Sony sensors such as in the D300 and D3x, the benefits of raising the ISO tail off somewhere between ISO 800 and 1600.

There are also a few footnotes to the above, such as the fact that there is almost nothing to be gained from using "intermediate" ISO's (the 1/3 stops in between ISO 100/200/400/800/1600 etc).

[Panopeeper]

for the 5D2, which has a lot of high ISO banding, it may make sense to continue raising the ISO to 3200 in step (3), since it lowers the banding noise relative to signal

Where do you see this happening?

[scubarob639]

Well people contradict their own beliefs in spite of truth all of the time. For instance, in the USA most people believe in the Constitutional phrase that "All men are created equal," which really means that all people deserve the same treatment. The same people who believed the equality phrase took until 1921 to see that women must be allowed to vote, or they were holding contradictory positions. The "Truth" I speak of here is not in the phrase itself, but is in the contradiction. Even though all contradictions are false (which is true), people still hold on to contradictory belief systems.

The point you make I understand, however the phrase "All men are created equal" is not in our constitution, and a women's right to vote is not an unalienable right as defined by our Declaration of Independence.

Rob

[ejmartin]

Where do you see this happening?

Since banding noise is a "feature" of the ISO amplifier/ADC, and thus not amplified when the ISO is increased, the signal is boosted relative to the banding noise by increasing the ISO.  The standard deviation of noise does not improve relative to signal beyond ISO 1600 in Canons, but on the other hand, banding noise is not a significant contributor to the standard deviation.  Better measures of the amount of banding noise involve either vertical/horizontal averages, which don't decrease the vertical/horizontal banding but average out the spatially random noise; or use of the Fourier transform of black frames to isolate the banding component.  I've not tested the 5D2, but I see this behavior on my 1D3.

[ejmartin]

#1 should be maximum aperture (lowest F-number)
#2 should be maximum shutter speed (the longer the better)

Yes on #1 -- minimum f-stop for required DOF.
On #2, the slowest exposure time in seconds that can freeze the action etc.

[Panopeeper]

Since banding noise is a "feature" of the ISO amplifier/ADC, and thus not amplified when the ISO is increased, the signal is boosted relative to the banding noise by increasing the ISO.

I guess any improvement should be apparent on the masked area, but I don't see that; the following captures show the masked area of the same sensor, with high contrast:

[dwdallam]

The point you make I understand, however the phrase "All men are created equal" is not in our constitution, and a women's right to vote is not an unalienable right as defined by our Declaration of Independence.

Rob

That is true enough. The point is that people believed in equality for all people, but denied females the same right they themselves enjoyed, which presents a contradiction. And you are right about unalienable rights too, but if you take your position to it's logical conclusion and run a country by it, you get a very Draconian State. But this is exactly what I am talking about: It takes some people a very long time to come to the conclusions that are not self contradictory. For instance, it is neither an unalienable right for us to talk freely on the internet. Take that to it logical conclusion and it would be just dandy for our government to censor this medium, just like China does for its populace. Having a right to be female and have an education isn't unalienable right either, and you get my point.

I think your comment that "A woman's right to vote is not an unalienable right" is a fact, but would create an extremely specious argument form.

[james_elliot]

I think this sentence is more or less the translation of the fundamental sentence of the  "Déclaration des droits de l'homme et du citoyen" of the french revolution.
The french sentence is very often translated in English by "All men are created free and equal" (the translation of "Tous les homme naissent libres et égaux"). However, the french sentence is indeed "Tous les hommes naissent libres et égaux en droit" which means "All men are created free and with equal rights".
However, women's right to vote in France was only established in 1944 (end of WWII, 150 years after the french revolution)... As Orwell's wrote: "all animals are equal but some animals are more equal than others" (Animal farm)

The point you make I understand, however the phrase "All men are created equal" is not in our constitution, and a women's right to vote is not an unalienable right as defined by our Declaration of Independence.

Rob

[marcmccalmont]

The basic rule of thumb in most situations is

1) Choose the minimum aperture for the required depth of field.
2) Choose the minimum shutter speed required to freeze the action, prevent camera shake, etc.
3) Choose the ISO such that the histogram is pushed to the right as much as possible while keeping the highlights that you want to preserve unclipped.

There is one major exception to this rule of thumb, when item (3) suggests an ISO higher than about 1600.  In this case, one is usually better off underexposing at ISO 1600 at the aperture and shutter speed suggested by (1) and (2), since there is little to no improvement in S/N above ISO on most current cameras.  The threshold for making such a decision varies from camera model to camera model; for the 5D2, which has a lot of high ISO banding, it may make sense to continue raising the ISO to 3200 in step (3), since it lowers the banding noise relative to signal, while for the new Sony sensors such as in the D300 and D3x, the benefits of raising the ISO tail off somewhere between ISO 800 and 1600.

There are also a few footnotes to the above, such as the fact that there is almost nothing to be gained from using "intermediate" ISO's (the 1/3 stops in between ISO 100/200/400/800/1600 etc).

Well I put your rule to the test with several low light exposures, 5DII,  color chart and measured the noise with noiseware professional (and my eye), the 100 ISO 1 minute exposure was much cleaner than the 4 second ISO 1600 exposure!
Marc

[ejmartin]

Well I put your rule to the test with several low light exposures, 5DII,  color chart and measured the noise with noiseware professional (and my eye), the 100 ISO 1 minute exposure was much cleaner than the 4 second ISO 1600 exposure!
Marc

Here's my test:  1D3 at ISO 100, 100% crop



and at ISO 1600 with the same Tv/Av:



the lower left square of the color chart is about 9 stops down from RAW saturation on the ISO 100 shot, IIRC.