High ISO noise and ADC

[FingerPainter]

There is an often repeated declaration that high ISO reduces noise. This is caused by the signal processing when the sensor image is processed and conditioned for the ADC. Although the reduction does happen, I have never seen this to be visually significant. You can infer this in D Cox thread a week ago With his SNR test using Photoshop to indicate the STD deviation of the test patches. Indeed a small improvement of SNR above ISO 1270 does happen.

I shot Canon 1D and 5D series extensively for low light performing arts between iso 3200 and ISO 12800 but have never seen a significant improvement of SNR. A number of years ago when this was first discussed, I did the appropriate test and, with pixel peeping, the noise reduction does exist but not to any degree to be of an advantage when shooting theatre or dance.

It would be nice to see a practical example of the noise reduction for real world shooting.

Am I misunderstanding or missing a salient point ?

How far is far?

If there is any difference in input-referred read noise at various ISO settings, whether or not you see it depends on exposure, subject matter, and light color, as well as display brightness of the shadow areas. When one of the three color channels has very weak sensor capture (think halogen for blue or deep forest shade for red), read noise becomes much stronger (relatively) in that channel, and is more likely to be visible.

Indeed those attributes can affect the appearance of noise. Would this be a strong difference in normal shooting at ISO 1600 to 12800 of normal scenes for performing arts?

A strong difference? No I don't think so. On most cameras, the effect is more pronounced for a three stop increase in ISO from a 100 or 200 ISO starting point than it is for a 3 stop increase at a higher starting ISO. I showed you the visible effect of a 4 stop increase from ISO 100, IDK if bobn2 still has that series of shots, and if so, how much higher the ISO goes in the series.

Does noise decrease with increasing ISO? As far as I can determine, a fractional decrease does happen but not to any photographically useful extent.

There is no doubt it is less useful at higher ISOs, and how useful depends on the particular camera. Look at Bill Claff's shadow improvement chart for your camera. That should give you an idea of what range of ISO increases is "useful".

 

[tony field]

There is an often repeated declaration that high ISO reduces noise. This is caused by the signal processing when the sensor image is processed and conditioned for the ADC. Although the reduction does happen, I have never seen this to be visually significant. You can infer this in D Cox thread a week ago With his SNR test using Photoshop to indicate the STD deviation of the test patches. Indeed a small improvement of SNR above ISO 1270 does happen.

I shot Canon 1D and 5D series extensively for low light performing arts between iso 3200 and ISO 12800 but have never seen a significant improvement of SNR. A number of years ago when this was first discussed, I did the appropriate test and, with pixel peeping, the noise reduction does exist but not to any degree to be of an advantage when shooting theatre or dance.

It would be nice to see a practical example of the noise reduction for real world shooting.

Am I misunderstanding or missing a salient point ?

How far is far?

If there is any difference in input-referred read noise at various ISO settings, whether or not you see it depends on exposure, subject matter, and light color, as well as display brightness of the shadow areas. When one of the three color channels has very weak sensor capture (think halogen for blue or deep forest shade for red), read noise becomes much stronger (relatively) in that channel, and is more likely to be visible.

Indeed those attributes can affect the appearance of noise. Would this be a strong difference in normal shooting at ISO 1600 to 12800 of normal scenes for performing arts?

A strong difference? No I don't think so. On most cameras, the effect is more pronounced for a three stop increase in ISO from a 100 or 200 ISO starting point than it is for a 3 stop increase at a higher starting ISO. I showed you the visible effect of a 4 stop increase from ISO 100, IDK if bobn2 still has that series of shots, and if so, how much higher the ISO goes in the series.

That is a highly constructed example to illustrate a technical ISO point. I seriously doubt that people would experience this in normal shooting

Does noise decrease with increasing ISO? As far as I can determine, a fractional decrease does happen but not to any photographically useful extent.

There is no doubt it is less useful at higher ISOs, and how useful depends on the particular camera. Look at Bill Claff's shadow improvement chart for your camera. That should give you an idea of what range of ISO increases is "useful".

Indeed Bill's chart can be used for this analysis

 

[John Sheehy]

There is no doubt it is less useful at higher ISOs, and how useful depends on the particular camera. Look at Bill Claff's shadow improvement chart for your camera. That should give you an idea of what range of ISO increases is "useful".

Indeed Bill's chart can be used for this analysis

Not as a full proof, though. That requires looking at the spatial character of noise, which can evade standard deviations. The post-gain read noise in the Nikon D5, the Canon 5D2, and some other cameras can have a lot of banding that doesn't register much in terms of standard deviation.

 

[ThrillaMozilla]

There is an often repeated declaration that high ISO reduces noise. This is caused by the signal processing when the sensor image is processed and conditioned for the ADC. Although the reduction does happen, I have never seen this to be visually significant. You can infer this in D Cox thread a week ago With his SNR test using Photoshop to indicate the STD deviation of the test patches. Indeed a small improvement of SNR above ISO 1270 does happen.

I shot Canon 1D and 5D series extensively for low light performing arts between iso 3200 and ISO 12800 but have never seen a significant improvement of SNR. A number of years ago when this was first discussed, I did the appropriate test and, with pixel peeping, the noise reduction does exist but not to any degree to be of an advantage when shooting theatre or dance.

It would be nice to see a practical example of the noise reduction for real world shooting.

Am I misunderstanding or missing a salient point ?

How far is far?

If there is any difference in input-referred read noise at various ISO settings, whether or not you see it depends on exposure, subject matter, and light color, as well as display brightness of the shadow areas. When one of the three color channels has very weak sensor capture (think halogen for blue or deep forest shade for red), read noise becomes much stronger (relatively) in that channel, and is more likely to be visible.

Indeed those attributes can affect the appearance of noise. Would this be a strong difference in normal shooting at ISO 1600 to 12800 of normal scenes for performing arts?

A strong difference? No I don't think so. On most cameras, the effect is more pronounced for a three stop increase in ISO from a 100 or 200 ISO starting point than it is for a 3 stop increase at a higher starting ISO. I showed you the visible effect of a 4 stop increase from ISO 100, IDK if bobn2 still has that series of shots, and if so, how much higher the ISO goes in the series.

That is a highly constructed example to illustrate a technical ISO point. I seriously doubt that people would experience this in normal shooting

They certainly can experience it. Try shooting a strongly backlit sample with an older Canon camera. You can easily see the read noise no matter what ISO setting you use. With a Nikon D7200 you should get much better results, and using a low ISO setting will get you a much higher dynamic range. It's a kludge, and one we shouldn't have to use, but it is a common situation.

Does noise decrease with increasing ISO? As far as I can determine, a fractional decrease does happen but not to any photographically useful extent.

There is no doubt it is less useful at higher ISOs, and how useful depends on the particular camera. Look at Bill Claff's shadow improvement chart for your camera. That should give you an idea of what range of ISO increases is "useful".

Indeed Bill's chart can be used for this analysis

 

[ThrillaMozilla]

That requires looking at the spatial character of noise, which can evade standard deviations. The post-gain read noise in the Nikon D5, the Canon 5D2, and some other cameras can have a lot of banding that doesn't register much in terms of standard deviation.

What?

Noise is almost always measured spatially. Banding certainly will be part of the standard deviation--unless one measures SD on a ridiculously small spot.

 

[FingerPainter]

.... Would this be a strong difference in normal shooting at ISO 1600 to 12800 of normal scenes for performing arts?

A strong difference? No I don't think so. On most cameras, the effect is more pronounced for a three stop increase in ISO from a 100 or 200 ISO starting point than it is for a 3 stop increase at a higher starting ISO. I showed you the visible effect of a 4 stop increase from ISO 100, IDK if bobn2 still has that series of shots, and if so, how much higher the ISO goes in the series.

That is a highly constructed example to illustrate a technical ISO point. I seriously doubt that people would experience this in normal shooting

It depends on what you shoot. It applies if you shoot JPEG in situations where you cannot get sufficient lightness at base ISO, and in RAW shooting where you cannot use up the highlight headroom a base ISO. Such situations occur when most often both your aperture and shutter settings are constrained, for instance as they may be with sports, and some events.

 

[tony field]

There is an often repeated declaration that high ISO reduces noise. This is caused by the signal processing when the sensor image is processed and conditioned for the ADC. Although the reduction does happen, I have never seen this to be visually significant. You can infer this in D Cox thread a week ago With his SNR test using Photoshop to indicate the STD deviation of the test patches. Indeed a small improvement of SNR above ISO 1270 does happen.

I shot Canon 1D and 5D series extensively for low light performing arts between iso 3200 and ISO 12800 but have never seen a significant improvement of SNR. A number of years ago when this was first discussed, I did the appropriate test and, with pixel peeping, the noise reduction does exist but not to any degree to be of an advantage when shooting theatre or dance.

It would be nice to see a practical example of the noise reduction for real world shooting.

Am I misunderstanding or missing a salient point ?

How far is far?

If there is any difference in input-referred read noise at various ISO settings, whether or not you see it depends on exposure, subject matter, and light color, as well as display brightness of the shadow areas. When one of the three color channels has very weak sensor capture (think halogen for blue or deep forest shade for red), read noise becomes much stronger (relatively) in that channel, and is more likely to be visible.

Indeed those attributes can affect the appearance of noise. Would this be a strong difference in normal shooting at ISO 1600 to 12800 of normal scenes for performing arts?

A strong difference? No I don't think so. On most cameras, the effect is more pronounced for a three stop increase in ISO from a 100 or 200 ISO starting point than it is for a 3 stop increase at a higher starting ISO. I showed you the visible effect of a 4 stop increase from ISO 100, IDK if bobn2 still has that series of shots, and if so, how much higher the ISO goes in the series.

That is a highly constructed example to illustrate a technical ISO point. I seriously doubt that people would experience this in normal shooting

They certainly can experience it. Try shooting a strongly backlit sample with an older Canon camera. You can easily see the read noise no matter what ISO setting you use. With a Nikon D7200 you should get much better results, and using a low ISO setting will get you a much higher dynamic range. It's a kludge, and one we shouldn't have to use, but it is a common situation.

Here is an old Canon 5D backlit scene at 100 ISO. It makes an excellent 16x20 print and a great magazine cover. I do not see any noticeable noise.



.

Higher 1600 ISO on the 5D has no practical noise or other issues imho. I had no issues shooting higher ISO up to 3200 max which is an extended ISO

Does noise decrease with increasing ISO? As far as I can determine, a fractional decrease does happen but not to any photographically useful extent.

There is no doubt it is less useful at higher ISOs, and how useful depends on the particular camera. Look at Bill Claff's shadow improvement chart for your camera. That should give you an idea of what range of ISO increases is "useful".

Indeed Bill's chart can be used for this analysis

--
Charles Darwin: "ignorance more frequently begets confidence than does knowledge."
tony

 

[John Sheehy]

That requires looking at the spatial character of noise, which can evade standard deviations. The post-gain read noise in the Nikon D5, the Canon 5D2, and some other cameras can have a lot of banding that doesn't register much in terms of standard deviation.

What?

Noise is almost always measured spatially. Banding certainly will be part of the standard deviation--unless one measures SD on a ridiculously small spot.

The standard deviation is completely independent of the spatial positions of the pixels in an image.

 

[tony field]

That requires looking at the spatial character of noise, which can evade standard deviations. The post-gain read noise in the Nikon D5, the Canon 5D2, and some other cameras can have a lot of banding that doesn't register much in terms of standard deviation.

What?

Noise is almost always measured spatially. Banding certainly will be part of the standard deviation--unless one measures SD on a ridiculously small spot.

The standard deviation is completely independent of the spatial positions of the pixels in an image.

Often true for sufficiently "large" sample areas

--
Charles Darwin: "ignorance more frequently begets confidence than does knowledge."
tony

 

[ThrillaMozilla]

Here is an old Canon 5D backlit scene at 100 ISO. It makes an excellent 16x20 print and a great magazine cover. I do not see any noticeable noise.

Here you go. Here's what happens when you try to see the backlit face. I have to admit, I've seen worse, but I've also seen more strongly backlit pictures too. It would probably look worse if it were bigger.







The sensor on the 5D ranks down there with my SL1, which DPR people like to make fun of.

 

[John Sheehy]

That requires looking at the spatial character of noise, which can evade standard deviations. The post-gain read noise in the Nikon D5, the Canon 5D2, and some other cameras can have a lot of banding that doesn't register much in terms of standard deviation.

What?

Noise is almost always measured spatially. Banding certainly will be part of the standard deviation--unless one measures SD on a ridiculously small spot.

The standard deviation is completely independent of the spatial positions of the pixels in an image.

Often true for sufficiently "large" sample areas.

Always true for an image. The spatial relationship of original pixels does not affect their standard deviation, ever.

Standard deviation means you take the population of samples, find their mean, find the difference between the mean and each sample, square those, add them, divide by the number of samples, and take the square root of that. Nothing in the definition makes any reference to which other samples each sample has for a neighbor.

Spatial relationships could affect the standard deviation of a binning or resampling of the original pixels. IOW, you could have two images with the same standard deviation, bin both 4x4, and get different standard deviations.

 

[tony field]

Here is an old Canon 5D backlit scene at 100 ISO. It makes an excellent 16x20 print and a great magazine cover. I do not see any noticeable noise.

Here you go. Here's what happens when you try to see the backlit face. I have to admit, I've seen worse, but I've also seen more strongly backlit pictures too. It would probably look worse if it were bigger.



The sensor on the 5D ranks down there with my SL1, which DPR people like to make fun of.

Your equestrian image is very strange. The 5D has poor (barely adequate) dynamic range and any shadow processing has to be done with care. You seem to have overdone shadow recovery. I have never had this degree of problem

--
Charles Darwin: "ignorance more frequently begets confidence than does knowledge."
tony

 

[ThrillaMozilla]

That requires looking at the spatial character of noise, which can evade standard deviations. The post-gain read noise in the Nikon D5, the Canon 5D2, and some other cameras can have a lot of banding that doesn't register much in terms of standard deviation.

What?

Noise is almost always measured spatially. Banding certainly will be part of the standard deviation--unless one measures SD on a ridiculously small spot.

The standard deviation is completely independent of the spatial positions of the pixels in an image.

What are you talking about? Noise is measured on images, in regions of interest with uniform exposure. Here's an example. The noise is the standard deviation of pixel values in the marked region interest. Ordinarily it would be measured on R, G, and B values separately, and usually on raw values. (Typically it would be measured on a step of a step wedge for better uniformity.)



Here's the histogram. The measured noise includes contributions from read noise, shot noise, and fixed pattern noise, including the banding. In this example the spatial characteristic of the noise in no way "evades" the standard deviation.

 

[ThrillaMozilla]

Standard deviation means you take the population of samples, find their mean, find the difference between the mean and each sample, square those, add them, divide by the number of samples, and take the square root of that.

Yes, that's (almost) a formula for calculating the standard difference. But in image processing and camera sensitometry we rarely measure it with sequential measurements.

It's actually done by using an image processor to measure the standard deviation in a region interest. Once you have defined the region of interest, it takes exactly one keystroke. An example precedes this.

Nothing in the definition makes any reference to which other samples each sample has for a neighbor.

Spatial relationships could affect the standard deviation of a binning or resampling of the original pixels. IOW, you could have two images with the same standard deviation, bin both 4x4, and get different standard deviations.

You can get more information if you also make sequential measurements, but I'm not sure how common that is. Are you trying to say that Claff's measurements are sequential only, and that he excludes spatial information from the noise data? That seems bizarre, but you'd have to ask him.

 

[ThrillaMozilla]

Your equestrian image is very strange. The 5D has poor (barely adequate) dynamic range and any shadow processing has to be done with care. You seem to have overdone shadow recovery. I have never had this degree of problem

Yes it is very strange. Shadow recovery is exactly what this is about, when you try to take an image of a backlit subject, if you try to show both subject and background nicely in the same photograph.

That's why you need low read noise. Some cameras would have done very much better than this.

 

[tony field]

Your equestrian image is very strange. The 5D has poor (barely adequate) dynamic range and any shadow processing has to be done with care. You seem to have overdone shadow recovery. I have never had this degree of problem

Yes it is very strange. Shadow recovery is exactly what this is about, when you try to take an image of a backlit subject, if you try to show both subject and background nicely in the same photograph.

That's why you need low read noise. Some cameras would have done very much better than this.

Low read noise is most useful as is dynamic range etc. I usually found that I can work within the technical limits of any digital camera I have owned and resulted in completely useful shots.

--
Charles Darwin: "ignorance more frequently begets confidence than does knowledge."
tony

 

[John Sheehy]

The standard deviation is completely independent of the spatial positions of the pixels in an image.

What are you talking about? Noise is measured on images, in regions of interest with uniform exposure. Here's an example. The noise is the standard deviation of pixel values in the marked region interest. Ordinarily it would be measured on R, G, and B values separately, and usually on raw values. (Typically it would be measured on a step of a step wedge for better uniformity.)



Here's the histogram. The measured noise includes contributions from read noise, shot noise, and fixed pattern noise, including the banding. In this example the spatial characteristic of the noise in no way "evades" the standard deviation.

What do you suppose would happen if you ran a macro that randomly swapped pixels around in the selection area? The standard deviation and the histogram would remain exactly the same, but the banding would be gone.

This is already demosaiced, though, which is really a poor context for this part of the discussion.

I don't know where the images are now, but several years ago I created synthetic raw blackframes that all had the same standard deviation, the same pixels (but in different locations), and the same histogram, which, when demosaiced, all looked completely different, with a huge difference in visible noise.

--
Beware of correct answers to wrong questions.
John

http://www.pbase.com/image/55384958.jpg

 

[ThrillaMozilla]

What do you suppose would happen if you ran a macro that randomly swapped pixels around in the selection area? The standard deviation and the histogram would remain exactly the same, but the banding would be gone.

Obviously, it tells you nothing about the spatial structure of the noise. Is that all you're saying? Gosh, the Internet can be a poor medium for exchange sometimes.

This is already demosaiced, though, which is really a poor context for this part of the discussion.

I don't know where the images are now, but several years ago I created synthetic raw blackframes that all had the same standard deviation, the same pixels (but in different locations), and the same histogram, which, when demosaiced, all looked completely different, with a huge difference in visible noise.

 

[fferreres]

Here is an old Canon 5D backlit scene at 100 ISO. It makes an excellent 16x20 print and a great magazine cover. I do not see any noticeable noise.

Here you go. Here's what happens when you try to see the backlit face. I have to admit, I've seen worse, but I've also seen more strongly backlit pictures too. It would probably look worse if it were bigger.



The sensor on the 5D ranks down there with my SL1, which DPR people like to make fun of.

It's not about being backlit. It's about being underexposed. If you can make up your mind before shooting, then the 5D still makes great images, and Noise Reduction has gone a long way since 2003 when the camera was launched. However, the sensor has 10 stops of DR, and read noise will start to become nasty if you could go to ISO 6400 or 12800 (which you can do by...underexposing on ISO 3200).

Do I like the low read noise of my A7 RIII more than the 5D? For sure. I think the original point was to say read noise is more important under severe low light, because the read noise can battle the signal.

 

[bobn2]

Standard deviation means you take the population of samples, find their mean, find the difference between the mean and each sample, square those, add them, divide by the number of samples, and take the square root of that.

Yes, that's (almost) a formula for calculating the standard difference. But in image processing and camera sensitometry we rarely measure it with sequential measurements.

Dimensions in space and time are equivalent. The difference between sampling over space is that we aren't actually using the same sensing device, but one we assume to be the same. If, in fact, they are not the same, pattern 'noise' is the result.