# A handy metric for "noise performance potential"

#### Doug Kerr

##### Active member
Often, a we contemplate buying/using one camera (and perhaps lens) vs. another, one consideration is what I will call (intentionally vaguely) potential noise performance.

Simplistically, we know that the greater the luminous energy on each pixel (for any given scene luminance, under "standard exposure" conditions), the better will be the noise performance. As a tactical matter, that means the higher the sensitivity we can employ and still get some certain image quality (from a noise standpoint).

With regard to how different camera-lens combinations will allow us to play this, two properties loom as important (in an "all other things being equal" context):

• The per-pixel area.

• The maximum aperture the lens offers (or the largest aperture we are interested in using based on depth of field and so forth).

We can combine those two into a simple "relative" metric for "potential noise performance" with this equation:
z = (p/N)^2​

where z is my potential noise performance metric, p is the pixel pitch, in μm, and N is the aperture, as an f-number.

It is interesting to reckon this metric for two of my current cameras, my Panasonic FZ1000 (with a "one inch" sensor and a pixel pitch of 2.41 μm) and my Canon G16 (with a "1/1.7 inch" sensor and a pixel pitch of 1.07 μm).

If we think in terms of operation at the smallest focal length of the lenses on both cameras, where the Z1000 has a maximum aperture of f/2.8 and the G16 has a maximum aperture of f/1.8, then the FZ1000 has z=0.7 and the G16 has z = 1.1, the G16 having a slightly greater potential noise performance.

If now we consider operation at 125 mm ff35e focal length, and again assume operation at the largest possible aperture, the FZ1000 shows z=0.38 and the G16 shows z=0.44. Again, the G16 shows a larger value of z than the FZ1000.

Of course, there is a difference in geometric resolution between these two cameras (with the FZ1000 being a 20 Mpx machine and the G16 12 Mpx), so we certainly can't conclude that the G16 is a better "low light" camera. But the comparison in terms of the metric I discuss here is interesting.

Now for a further comparison, for a Canon EOS M50, equipped with the infamous EF-M 22 mm f/2.0 lens (ff35e 35 mm), we get z=3.5, suggesting (at that focal length!) a superior noise performance potential.

Of course, the reason I include the modifier "potential" here is that this is a rather naïve estimation of noise performance. But I think useful.

Best regards,

Doug

#### James Lemon

##### Active member
Often, a we contemplate buying/using one camera (and perhaps lens) vs. another, one consideration is what I will call (intentionally vaguely) potential noise performance.

Simplistically, we know that the greater the luminous energy on each pixel (for any given scene luminance, under "standard exposure" conditions), the better will be the noise performance. As a tactical matter, that means the higher the sensitivity we can employ and still get some certain image quality (from a noise standpoint).

With regard to how different camera-lens combinations will allow us to play this, two properties loom as important (in an "all other things being equal" context):

• The per-pixel area.

• The maximum aperture the lens offers (or the largest aperture we are interested in using based on depth of field and so forth).

We can combine those two into a simple "relative" metric for "potential noise performance" with this equation:
z = (p/N)^2​

where z is my potential noise performance metric, p is the pixel pitch, in μm, and N is the aperture, as an f-number.

It is interesting to reckon this metric for two of my current cameras, my Panasonic FZ1000 (with a "one inch" sensor and a pixel pitch of 2.41 μm) and my Canon G16 (with a "1/1.7 inch" sensor and a pixel pitch of 1.07 μm).

If we think in terms of operation at the smallest focal length of the lenses on both cameras, where the Z1000 has a maximum aperture of f/2.8 and the G16 has a maximum aperture of f/1.8, then the FZ1000 has z=0.7 and the G16 has z = 1.1, the G16 having a slightly greater potential noise performance.

If now we consider operation at 125 mm ff35e focal length, and again assume operation at the largest possible aperture, the FZ1000 shows z=0.38 and the G16 shows z=0.44. Again, the G16 shows a larger value of z than the FZ1000.

Of course, there is a difference in geometric resolution between these two cameras (with the FZ1000 being a 20 Mpx machine and the G16 12 Mpx), so we certainly can't conclude that the G16 is a better "low light" camera. But the comparison in terms of the metric I discuss here is interesting.

Now for a further comparison, for a Canon EOS M50, equipped with the infamous EF-M 22 mm f/2.0 lens (ff35e 35 mm), we get z=3.5, suggesting (at that focal length!) a superior noise performance potential.

Of course, the reason I include the modifier "potential" here is that this is a rather naïve estimation of noise performance. But I think useful.

Best regards,

Doug
You may find this new sensor technology interesting Doug?

https://unews.utah.edu/bright-idea-for-lowlight-photography/

Best, regards
James

#### Asher Kelman

##### OPF Owner/Editor-in-Chief
Doug,

I am so pleased to read your new work, especially as I can tell my wife how wise my decision was to pay 10 times what is needed for good pictures in justifying my Fuji GFX!

I now know why ISO 6400 on my camera is so much like ISO 400 on my Canon 5D mark II.

Asher

#### Doug Kerr

##### Active member
Hi, Asher,

Doug,

I am so pleased to read your new work, especially as I can tell my wife how wise my decision was to pay 10 times what is needed for good pictures in justifying my Fuji GFX!

I now know why ISO 6400 on my camera is so much like ISO 400 on my Canon 5D mark II.
Well, for the Fuji GFX 50S, p=5.1 μm, while for the Canon EOS 5D Mark II p = 6.2 μm.

So we have an observed improvement factor of about 23.6:1 beyond what we might project from pixel size alone that results in some way not known to me.

And that kind of magic is certainly worth the cost!

Best regards,

Doug

#### Doug Kerr

##### Active member
I have been "reminded" by a skilled and experienced photographer that, if we contemplate working with the JPEG files from the cameras of interest rather than the raw files, it is foolish to be at all interested in pixel size.

Best regards,

Doug

#### Doug Kerr

##### Active member
To illuminate the implications of the notion I put forth at the beginning of this thread:

If we consider two cameras, each shooting at the maximum available aperture of the lens on board, the factor z will very broadly tell us, for a given noise impact, at what shutter speed we can shoot.

The chain here is that the pixel area gives us (very broadly) an indication of the ISO sensitivity we can employ while enjoying some certain noise performance, and then the basic exposure equation, taking into account that sensitivity, the available aperture, and the scene luminance, will tell us at what shutter speed we can shoot.

Best regards,

Doug

#### Asher Kelman

##### OPF Owner/Editor-in-Chief
Another 3 factors to consider are

1. back illumination which increases the efficiency of counting electric charge

2. shielding of pixels from electron spill from saturated sensels by space separation, (feasible in larger sensors for the same MP count) and electrical isolation.

3.Magic

#### Doug Kerr

##### Active member
Hi, Asher,

Another 3 factors to consider are

1. back illumination which increases the efficiency of counting electric charge

2. shielding of pixels from electron spill from saturated sensels by space separation, (feasible in larger sensors for the same MP count) and electrical isolation.

3.Magic
Well said.

So perhaps rather than making a SWAG (scientific wild-ass guess) we should look to some objective tests of sensor performance.

Of course one of the important accrediting agencies in that regard is DxOMark. An important parameter of their testing of sensors is the signal-to-noise ratio. But, for ease in comprehension by their constituency, they report this in terms of the greatest ISO sensitivity that can be engaged and still meet a certain signal-to-noise ratio criterion (a rather stringent one - 30 dB).

Here, I will list the DxOMark "total sensor score" (in"points") as well as their "low-light performance" metric (as ISO sensitivity) for several cameras. I use the forum's "code" rubric to allow me to make a readable table.

Code:
``````                   Total Low
Camera             score light

Canon 1Ds Mark III   80  1663
Canon 5DS R          86  2308
Canon G16            54   230
Canon G1X Mark II    58   581
Panasonic FZ1000     64   517
Panasonic ZS100      70   559
Sony A7R             95  2746``````
Sorry, the Fujifilm GFX 50S has not yet been tested.

Best regards,

Doug

#### Asher Kelman

##### OPF Owner/Editor-in-Chief
Hi, Asher,

Well said.

So perhaps rather than making a SWAG (scientific wild-ass guess) we should look to some objective tests of sensor performance.

Of course one of the important accrediting agencies in that regard is DxOMark. An important parameter of their testing of sensors is the signal-to-noise ratio. But, for ease in comprehension by their constituency, they report this in terms of the greatest ISO sensitivity that can be engaged and still meet a certain signal-to-noise ratio criterion (a rather stringent one - 30 dB).

Here, I will list the DxOMark "total sensor score" (in"points") as well as their "low-light performance" metric (as ISO sensitivity) for several cameras. I use the forum's "code" rubric to allow me to make a readable table.

Code:
``````                   Total Low
Camera             score light

Canon 1Ds Mark III   80  1663
Canon 5DS R          86  2308
Canon G16            54   230
Canon G1X Mark II    58   581
Panasonic FZ1000     64   517
Panasonic ZS100      70   559
Sony A7R             95  2746``````
Sorry, the Fujifilm GFX 50S has not yet been tested.

Doug,

Kudos for the table and its formatting. Hope you will send me the secrets!

For the GFX we could use a blend of the Phase one or Pentax 50 MP as its pretty close to the same sensor, in the order of newness, Pentax, Phase One...then GFX.

We could perhaps see if there is any insight in the value of looking here at pixel size.

Asher

#### Asher Kelman

##### OPF Owner/Editor-in-Chief
Doug,

Two extra points.

1. Look at the beautiful night pictures with a modest size micro 4/3 here

2. When I disparage use of pixel size in evaluating “out-of-the-camera” jpg noise, it’s only because the MFRs have already applied noise removal/suppression algorithms that is akin to spiking cookies with entirely different illicit drugs and then asking which makes you happier!

Pixel size is important, but so is the structure of the sensel and magic factors, but at least with RAW files, the noise removal is “relatively” minor!

Asher

#### James Lemon

##### Active member
Keep in mind that using automatic modes will effect the signal noise ratio as well but using manual will help .

#### James Lemon

##### Active member
Hi, Asher,

Well said.

So perhaps rather than making a SWAG (scientific wild-ass guess) we should look to some objective tests of sensor performance.

Of course one of the important accrediting agencies in that regard is DxOMark. An important parameter of their testing of sensors is the signal-to-noise ratio. But, for ease in comprehension by their constituency, they report this in terms of the greatest ISO sensitivity that can be engaged and still meet a certain signal-to-noise ratio criterion (a rather stringent one - 30 dB).

Here, I will list the DxOMark "total sensor score" (in"points") as well as their "low-light performance" metric (as ISO sensitivity) for several cameras. I use the forum's "code" rubric to allow me to make a readable table.

Code:
``````                   Total Low
Camera             score light

Canon 1Ds Mark III   80  1663
Canon 5DS R          86  2308
Canon G16            54   230
Canon G1X Mark II    58   581
Panasonic FZ1000     64   517
Panasonic ZS100      70   559
Sony A7R             95  2746``````
Sorry, the Fujifilm GFX 50S has not yet been tested.

Best regards,

Doug
Hello Doug

Sensor characteristics such as actual sensor size, sensor area, pixel pitch, pixel area, pixel density, crop factor, and more.

You may find this camera data base useful in your research but no medium format.

https://www.digicamdb.com/

Best, regards
James

#### Doug Kerr

##### Active member
Hi, James,

Hello Doug

Sensor characteristics such as actual sensor size, sensor area, pixel pitch, pixel area, pixel density, crop factor, and more.

You may find this camera data base useful in your research but no medium format.

https://www.digicamdb.com/
Thank you for the link to that very valuable resource. I had not been aware of it before.

Best regards,

Doug

#### Doug Kerr

##### Active member
Hi, Asher,

Doug,

2. When I disparage use of pixel size in evaluating “out-of-the-camera” jpg noise, it’s only because the MFRs have already applied noise removal/suppression algorithms that is akin to spiking cookies with entirely different illicit drugs and then asking which makes you happier!
Sure. But there are always tradeoffs, such as loss of detail as more noise is removed (and the camera reviews often discuss and demonstrate this at length). So we pay one way or the other for the original sin of noise in the sensor output. Whether some metric of "overall negative impact of sensor noise on the 'quality' of the delivered photograph" tracks well with pixel size I don't know.

Pixel size is important, but so is the structure of the sensel and magic factors, but at least with RAW files, the noise removal is “relatively” minor!
Then, we apply noise reduction one way or another in the postprocessing, where we may have tools that "do better for us" than those included in the camera. Or, for our basic "development" tool (especially if it is provided by the camera manufacturer, as for Canon's DPP), we may have access to essentially the same adjustable tool used in the camera during the preparation of the JPEG output.

Best regards,

Doug

#### Asher Kelman

##### OPF Owner/Editor-in-Chief
The corresponding scientific article is here: https://www.osapublishing.org/optica/abstract.cfm?uri=optica-2-11-933 and explains the method actually used. The proposed method needs 9 times the number of pixels as a standard Bayer array. It was proposed in 2015 (so not that "new"...), I am not surprised that it has not seen any commercial implementation.
But one of the methods will eventually be inexpensive to mass produce.

BTW, where does the main components of noise come from, the sensel well or the A/D converter and electronics behind it?

Asher

#### Jerome Marot

##### Active member
BTW, where does the main components of noise come from, the sensel well or the A/D converter and electronics behind it?
In modern sensors and in low light situations, the main component of noise is the quantum nature of light: photon shot noise.

#### Asher Kelman

##### OPF Owner/Editor-in-Chief
In modern sensors and in low light situations, the main component of noise is the quantum nature of light: photon shot noise.
So would that be [square root of n]/n just like counting photons in a liquid scintillation counter?

But where do we get the effect of total sensel size?

Or does that cancel out as the flux/unit area will not change?

Asher

#### Jerome Marot

##### Active member
I am not sure I understand your question. Sensel size is directly linked to the number of photons?

#### Doug Kerr

##### Active member
Hi, Asher,

So would that be [square root of n]/n just like counting photons in a liquid scintillation counter?
Yes. And, for future reference, note that sqrt/n= 1/sqrt.

But where do we get the effect of total sensel size?

Or does that cancel out as the flux/unit area will not change?
For operation at a given ISO sensitivity and for a given given scene, the exposure ordained by a "standard" metering system will have a certain phtometric exposure, H, on the sensor for any given small region of the image. H is the product of the illuminance on the sensor (luminous flux per unit area) and the exposure time. That will be the same for any overall size sensor, or any sensel pitch.

Given that, the expected (average) total number of photons incident on each sensel (in that region of the image) during that exposure time will be directly proportional to the effective collection area of the sensel (which we often approximate by the square of the sensel pitch).

Since the arrival of the photons at the sensel is essentially a Poisson process, the standard deviation of the photons' impact on the sensel during the exposure (and thus of the output voltage), the indicator of noise-to-signal ratio, will be less. Simplistically, the signal-to-noise ratio is the square root of the expected number of photons.

Best regards,

Doug

#### James Lemon

##### Active member
The corresponding scientific article is here: https://www.osapublishing.org/optica/abstract.cfm?uri=optica-2-11-933 and explains the method actually used. The proposed method needs 9 times the number of pixels as a standard Bayer array. It was proposed in 2015 (so not that "new"...), I am not surprised that it has not seen any commercial implementation.
Jerome

Most of these designs aim to enhance the color accuracy by tuning the transmitted spectral bands via nano- or microstructures. Plasmonics-based color filters suffer from decreased light transmission due to parasitic absorption of the required metal layers [4–10]. In addition, these require very precise nanofabrication of subwavelength structures, which can be challenging and experience difficulties in extension to mass production. Alternative filters that utilize a variety of optical resonance effects have also been proposed.

The innovative process can take a long time but once they overcome the hurdles things can take off exponentially. It can take longer to go from .003% to 1% than it does to go from 1% to 90%, in the development, this would be my guess?

Best, regards
James

#### Asher Kelman

##### OPF Owner/Editor-in-Chief
Jerome

Most of these designs aim to enhance the color accuracy by tuning the transmitted spectral bands via nano- or microstructures. Plasmonics-based color filters suffer from decreased light transmission due to parasitic absorption of the required metal layers [4–10]. In addition, these require very precise nanofabrication of subwavelength structures, which can be challenging and experience difficulties in extension to mass production. Alternative filters that utilize a variety of optical resonance effects have also been proposed.

The innovative process can take a long time but once they overcome the hurdles things can take off exponentially. It can take longer to go from .003% to 1% than it does to go from 1% to 90%, in the development, this would be my guess?

Best, regards
James
There are a lot of extra parameters like degree of axial rotation. I wonder whether that can be used too? Beam splitters seem to have passed entirel!

Asher

#### Ted Cousins

##### New member
The Foveon Shows Promise ...

Often, a we contemplate buying/using one camera (and perhaps lens) vs. another, one consideration is what I will call (intentionally vaguely) potential noise performance.

Simplistically, we know that the greater the luminous energy on each pixel (for any given scene luminance, under "standard exposure" conditions), the better will be the noise performance. As a tactical matter, that means the higher the sensitivity we can employ and still get some certain image quality (from a noise standpoint).

With regard to how different camera-lens combinations will allow us to play this, two properties loom as important (in an "all other things being equal" context):

• The per-pixel area.

• The maximum aperture the lens offers (or the largest aperture we are interested in using based on depth of field and so forth).

We can combine those two into a simple "relative" metric for "potential noise performance" with this equation:
z = (p/N)^2​

where z is my potential noise performance metric, p is the pixel pitch, in μm, and N is the aperture, as an f-number.

It is interesting to reckon this metric for two of my current cameras, my Panasonic FZ1000 (with a "one inch" sensor and a pixel pitch of 2.41 μm) and my Canon G16 (with a "1/1.7 inch" sensor and a pixel pitch of 1.07 μm).

If we think in terms of operation at the smallest focal length of the lenses on both cameras, where the Z1000 has a maximum aperture of f/2.8 and the G16 has a maximum aperture of f/1.8, then the FZ1000 has z=0.7 and the G16 has z = 1.1, the G16 having a slightly greater potential noise performance.

If now we consider operation at 125 mm ff35e focal length, and again assume operation at the largest possible aperture, the FZ1000 shows z=0.38 and the G16 shows z=0.44. Again, the G16 shows a larger value of z than the FZ1000.

Of course, there is a difference in geometric resolution between these two cameras (with the FZ1000 being a 20 Mpx machine and the G16 12 Mpx), so we certainly can't conclude that the G16 is a better "low light" camera. But the comparison in terms of the metric I discuss here is interesting.

Now for a further comparison, for a Canon EOS M50, equipped with the infamous EF-M 22 mm f/2.0 lens (ff35e 35 mm), we get z=3.5, suggesting (at that focal length!) a superior noise performance potential.

Of course, the reason I include the modifier "potential" here is that this is a rather naïve estimation of noise performance. But I think useful.

Best regards,

Doug
An interesting metric, Doug, which I belatedly put to the test on an early Sigma (SD9) camera:

z = (9.12/2.8)^2 = a whopping noise performance potential of 10.61 ...

... in reality, of course, that gets totally destroyed during conversion from raw to RGB to the extent that any ISO > 200 is a waste of time unless you happen to like color-blotching, LOL.

#### Asher Kelman

##### OPF Owner/Editor-in-Chief
An interesting metric, Doug, which I belatedly put to the test on an early Sigma (SD9) camera:

z = (9.12/2.8)^2 = a whopping noise performance potential of 10.61 ...

... in reality, of course, that gets totally destroyed during conversion from raw to RGB to the extent that any ISO > 200 is a waste of time unless you happen to like color-blotching, LOL.
Ted,

Could you explain that 10.61 in real world examples?

I would mind if I could only do magic at ISO 200!

Asher