Are bigger pixels less noisy?

Started 5 months ago | Discussions
alanr0 Senior Member • Posts: 2,472
Re: Noise reduction

JahnG wrote:

Alan, interesting issues concerning noise reduction.

My original pixel-size and noise questions have been thoroughly and very well answered, so yes, why not slightly look at noise reduction, which probably is quite interesting for many.

Not something I know a great deal about.

I did get some satisfying results using Neat Image, but that was 10 years ago, and things have moved on since then.

Have you checked out the DPR Retouching Forum. There are some active threads on noise reduction there, including some specifically about Topaz Denoise which J A C S mentioned.

I shoot jpg and very seldom use NR in my old PaintShop Pro 9 program, which mainly seems to just blur the picture when using NR. Only when there has been some easily defined ("lassoed") noisy area on a picture I have "blurred" that area. I do own a "newer" PaintShop Pro program, the X5 version, which I have seldom used as the old version "9" works perfectly well in Win10 pro and is slightly simpler and faster to use. Both programs basically seems to do the same thing.

I just checked (in the net) that my newer X5 version has an additional NR setting, "digital noise" where different colored areas might be "noise reduced". Must look at it, but I doubt it is very advanced?

However, it seems that some more advanced noise reduction plugins would be possible to utilize not only in Photoshop but also in PaintShop Pro X5. (I'm Shooting jpg)

I found a program in Amazon.de (which I use, being in EU) named AKVIS Noise Buster, but it is probably not anything advanced at all? Any other suggestions?

What are BTW "non linear" noise reduction systems, some plugins?

One particularly simple noise reduction scheme is a linear low-pass filter. This attenuates high spatial frequencies. Essentially it averages out the noise, which it assumes is random. Unfortunately, this also smears out wanted detail in the image. A variation on this is to smooth out low amplitude noise below some threshold. Larger changes are assumed to correspond to actual wanted detail, and are passed through mostly unaltered. This is one version of a non-linear process, and thresholding is a common option in image processing packages.

Another non-linear technique is median filtering. Instead of finding the average intensity in a 5 or 9 pixel square, one sorts the nearest neighbour pixels in order of size, and replaces the central pixel with the intensity from the middle of the list (the median value). This is an effective way to remove the "hot pixels" that J A C S and Eric Fossum discussed, but in its simplest form it too can lose al lot of wanted detail.

The most effective noise reduction algorithms have more sophisticated means to figure out which parts of the image are actual detail and what is noise. I am not up to speed on the latest developments here.

If you don't get a detailed answer in this thread, you could try asking a new question specifically about noise reduction. Some of the local signal processing experts could be avoiding this thread, as pixel size has been argued over repeatedly in the past.

In addition, you should find well informed voices in the the Retouching form, particularly on the subjective merits of different products.

Good luck!

-- hide signature --

Alan Robinson

JahnG
OP JahnG Veteran Member • Posts: 3,339
Re: Noise reduction

alanr0 wrote:

JahnG wrote:

Alan, interesting issues concerning noise reduction.

My original pixel-size and noise questions have been thoroughly and very well answered, so yes, why not slightly look at noise reduction, which probably is quite interesting for many.

Not something I know a great deal about.

I did get some satisfying results using Neat Image, but that was 10 years ago, and things have moved on since then.

Have you checked out the DPR Retouching Forum. There are some active threads on noise reduction there, including some specifically about Topaz Denoise which J A C S mentioned.

I shoot jpg and very seldom use NR in my old PaintShop Pro 9 program, which mainly seems to just blur the picture when using NR. Only when there has been some easily defined ("lassoed") noisy area on a picture I have "blurred" that area. I do own a "newer" PaintShop Pro program, the X5 version, which I have seldom used as the old version "9" works perfectly well in Win10 pro and is slightly simpler and faster to use. Both programs basically seems to do the same thing.

I just checked (in the net) that my newer X5 version has an additional NR setting, "digital noise" where different colored areas might be "noise reduced". Must look at it, but I doubt it is very advanced?

However, it seems that some more advanced noise reduction plugins would be possible to utilize not only in Photoshop but also in PaintShop Pro X5. (I'm Shooting jpg)

I found a program in Amazon.de (which I use, being in EU) named AKVIS Noise Buster, but it is probably not anything advanced at all? Any other suggestions?

What are BTW "non linear" noise reduction systems, some plugins?

One particularly simple noise reduction scheme is a linear low-pass filter. This attenuates high spatial frequencies. Essentially it averages out the noise, which it assumes is random. Unfortunately, this also smears out wanted detail in the image. A variation on this is to smooth out low amplitude noise below some threshold. Larger changes are assumed to correspond to actual wanted detail, and are passed through mostly unaltered. This is one version of a non-linear process, and thresholding is a common option in image processing packages.

Another non-linear technique is median filtering. Instead of finding the average intensity in a 5 or 9 pixel square, one sorts the nearest neighbour pixels in order of size, and replaces the central pixel with the intensity from the middle of the list (the median value). This is an effective way to remove the "hot pixels" that J A C S and Eric Fossum discussed, but in its simplest form it too can lose al lot of wanted detail.

The most effective noise reduction algorithms have more sophisticated means to figure out which parts of the image are actual detail and what is noise. I am not up to speed on the latest developments here.

If you don't get a detailed answer in this thread, you could try asking a new question specifically about noise reduction. Some of the local signal processing experts could be avoiding this thread, as pixel size has been argued over repeatedly in the past.

In addition, you should find well informed voices in the the Retouching form, particularly on the subjective merits of different products.

Good luck!

Thanks, will try

Jahn

 JahnG's gear list:JahnG's gear list
Canon PowerShot A650 IS Panasonic ZS100 Pentax K-r Pentax K100D Super Pentax K-S1
JahnG
OP JahnG Veteran Member • Posts: 3,339
Re: Topic changed to noise reduction

J A C S wrote:

Since you shoot JPEG, I think Topaz would be your best choice. I personally did not like it much but that was "a version ago", I have not tried the most recent one. It could not remove some hot pixels, for example.

If you shoot RAW, IMO the best choice is DXO. The downside is that it does not have the colors and the overall behavior of LR/ACR. It is possible to force it to work with a custom color profile but the shadow lifting and the highlight compression algorithms are not as good (IMO) as those of LR.

Thank you, I'll look at Topaz

Jahn

 JahnG's gear list:JahnG's gear list
Canon PowerShot A650 IS Panasonic ZS100 Pentax K-r Pentax K100D Super Pentax K-S1
J A C S
J A C S Forum Pro • Posts: 19,277
Re: Noise reduction
2

I know a bit about noise even though this is not in my main area of research.

First, it is good to know something about the statistical characteristics of the noise. Is it an additive, or not, etc.; then what is the distribution, the strength, and the auto-correlation, which also determines the spectrum. A typical example is to have white additive noise, which can describe, I guess, most of the read noise (but not the shot noise).

Without additional assumptions, you cannot remove noise, period. After all, I can take a perfect photo of a noisy gray patch, and all NR software will remove most of it but that would be actual signal.

Fortunately, most of our photos are not like that, but some parts could look close, like the sand on a beach, etc. Then one tries to guess what is noise and what is signal because signal is expected to be more orderly. Just applying a low-pass filter is not a great NR since white noise, for example, has uniform spectrum on average, and the medium and the low frequency noise remain - they look like grain. You kill fine detail and fine noise but you leave "coarse" noise.

Algorithms based on the assumption that the image is "orderly" (sparse is some basis) might be classified as compressed sensing in a wide sense (I include various optimizations here, TV included). They work when they do and fail when they fail; for example, TV kills patterns. AFAIK, there are other algorithms looking for edges for example (the old DXO Prime claimed that) but I do not know details.

The newest trend is machine learning (AI). It tries to match parts of the image with parts of images in the "training set" somehow. There are lot of parameters to adjust there, there is a lot of trial and error and little theoretical understanding but they "work". DXO Prime AI, for example, does wonders IMO. It works on RAW only. This is what separates photography from other applications - we have three channels while many of the traditional techniques assume one.

Iliah Borg Forum Pro • Posts: 28,665
Re: Quantum efficiency, 93° ;)
5

Nick Zochios wrote:

Ohhhhh I see.
Sorry, but I will destroy your dream.

But you can't destroy Ούζο Πλωμάρι

-- hide signature --
JahnG
OP JahnG Veteran Member • Posts: 3,339
Re: Noise reduction

J A C S wrote:

I know a bit about noise even though this is not in my main area of research.

First, it is good to know something about the statistical characteristics of the noise. Is it an additive, or not, etc.; then what is the distribution, the strength, and the auto-correlation, which also determines the spectrum. A typical example is to have white additive noise, which can describe, I guess, most of the read noise (but not the shot noise).

Without additional assumptions, you cannot remove noise, period. After all, I can take a perfect photo of a noisy gray patch, and all NR software will remove most of it but that would be actual signal.

Fortunately, most of our photos are not like that, but some parts could look close, like the sand on a beach, etc. Then one tries to guess what is noise and what is signal because signal is expected to be more orderly. Just applying a low-pass filter is not a great NR since white noise, for example, has uniform spectrum on average, and the medium and the low frequency noise remain - they look like grain. You kill fine detail and fine noise but you leave "coarse" noise.

Algorithms based on the assumption that the image is "orderly" (sparse is some basis) might be classified as compressed sensing in a wide sense (I include various optimizations here, TV included). They work when they do and fail when they fail; for example, TV kills patterns. AFAIK, there are other algorithms looking for edges for example (the old DXO Prime claimed that) but I do not know details.

The newest trend is machine learning (AI). It tries to match parts of the image with parts of images in the "training set" somehow. There are lot of parameters to adjust there, there is a lot of trial and error and little theoretical understanding but they "work". DXO Prime AI, for example, does wonders IMO. It works on RAW only. This is what separates photography from other applications - we have three channels while many of the traditional techniques assume one.

Thanks, so perhaps not to generally expect wonders.

Just looked at Topaz DeNoise comments, and some were not necessarily satisfied with the newest version(s)

Jahn

 JahnG's gear list:JahnG's gear list
Canon PowerShot A650 IS Panasonic ZS100 Pentax K-r Pentax K100D Super Pentax K-S1
J A C S
J A C S Forum Pro • Posts: 19,277
Re: Noise reduction

JahnG wrote:

Thanks, so perhaps not to generally expect wonders.

Just looked at Topaz DeNoise comments, and some were not necessarily satisfied with the newest version(s)

Jahn

You should download the trial version.

photonut2008
photonut2008 Veteran Member • Posts: 6,371
Re: Quantum efficiency and signal to noise ratio

Nick Zochios wrote:

I don't know if that says anything but this is a test shot with my D700 (12,1mpx) at ISO 8000!! No editing. The lens is the old Nikon 28-105 AF-D f/3,5-4,5
Just RAW to jpeg and then an automatic noise reduction through "DeNoise Ai" software.
It looks to me like an iso 400 shot. Very clear and full of details.

Nikon D700-iso 8000

Full of artifacts (false details) too.

ISO 6400 pushed .33 of a stop. Bottom crop is resized to match the D700 output and then post-processed using Topaz Denoise and Sharpening.

I took a lighter approach to the noise reduction to avoid the artifacting that is evident in the text of your bottle.

-- hide signature --

DPR, where gear is king and photography merely a jester

 photonut2008's gear list:photonut2008's gear list
Nikon D800 Nikon D500 Nikon AF-S Nikkor 70-200mm f/2.8G ED VR Nikon AF-S Teleconverter TC-14E II Nikon AF Nikkor 105mm f/2D DC +16 more
(unknown member) Forum Member • Posts: 61
Re: Quantum efficiency and signal to noise ratio

Yeap! Please use the old 80 bucks lens (28-105)  that i used for my own shot and then come back to tell us what you found.;-)

J A C S
J A C S Forum Pro • Posts: 19,277
Re: Quantum efficiency and signal to noise ratio
3

Nick Zochios wrote:

Yeap! Please use the old 80 bucks lens (28-105) that i used for my own shot and then come back to tell us what you found.;-)

Why are you so proud of your shot? It is awful.

bobn2
bobn2 Forum Pro • Posts: 69,811
Re: Are bigger pixels less noisy?

J A C S wrote:

JahnG wrote:

Iliah Borg wrote:

JahnG wrote:

Camera manufactors like Canon, Sony, (Panasonic) have made ”low noise” cameras with less and bigger ”pixels”. Do the manufactors not know what they are doing? (Some people in DPR say that such cameras are manufactored because the customers expect that sensors with bigger ””pixels” would be less noisy?)

Suppose 1 large pixel collects the same light as four smaller pixels occupying the same space. Is it possible?

Sounds logical.

The combined light collecting area of a FF sensor might (roughly) be the same, regardless of if there are 20M bigger or 50M smaller pixels.

But how about the combined read noise of 50M pixels compared to the combined read noise of 20M pixels. One might thus think that the sensor having only 40% pixel count would have much less combined read noise? (Or shouldn't we sum read noise?).

The engineers may chime in why - but what I see is that smaller pixels have smaller read noise. In the end, read noise per unit area seems weakly dependent on pixel size.

I had a huge great argument here with Eric on this topic some years ago. Anyhow, this is what I was saying, reworded a bit, in deference to Eric.

When you're looking for dependencies, you end up making some assumptions, which are likely not realistic, but if the reality carries some correlation with the assumptions, maybe you've captured 'weak dependency'.

So, let's assume that all pixels are designed the same (of course, they aren't) and we produce small pixels just by scaling large pixels down proportionately. This means that every feature of the pixel gets smaller (and also, another discussion, that it's fill factor stays the same). The input referred read noise depends on the electronic noise from the pixel source follower and downstream and the conversion gain of the pixel. In turn the conversion gain is inversely proportional to the capacitance of the SF gate and floating diffusion. If these are scaled down the capacitance goes down, the CG goes up and the read noise is reduced.

Now, as said before, pixels aren't in reality designed like that, but we can see why there might be a correlation. For instance:

Companies that have multiple fab lines will tend to use the smaller process nodes for smaller pixel sensors, so the features in those pixel designs will be smaller.

If the sensor is being able to cope with a particular saturation exposure (e.g. to allow use with 100 ISO) then at that exposure the smaller pixel collects fewer photoelectrons, and can be designed with a lower saturation capacity. That means smaller capacitance in the SF gate/floating diffusion.

-- hide signature --

Is it always wrong
for one to have the hots for
Comrade Kim Yo Jong?

bobn2
bobn2 Forum Pro • Posts: 69,811
Re: Are recent sensors less noisy?

alanr0 wrote:

The sensorgen site is no longer maintained, but the information extracted from DxO is still available on WaybackMachine.

Sorry about that - DxOMark kept of changing the data formats, and at the time I was too busy to spend the time needed to keep up with them, and in any case...

Bill Claff includes results for recent cameras on his photonstophotos site.

Bill was doing much the same job and his data is generally more reliable, so there didn't seem much point.

-- hide signature --

Is it always wrong
for one to have the hots for
Comrade Kim Yo Jong?

bodeswell Regular Member • Posts: 382
Re: Are bigger pixels less noisy?

bobn2 wrote:

J A C S wrote:

JahnG wrote:

Iliah Borg wrote:

JahnG wrote:

Camera manufactors like Canon, Sony, (Panasonic) have made ”low noise” cameras with less and bigger ”pixels”. Do the manufactors not know what they are doing? (Some people in DPR say that such cameras are manufactored because the customers expect that sensors with bigger ””pixels” would be less noisy?)

Suppose 1 large pixel collects the same light as four smaller pixels occupying the same space. Is it possible?

Sounds logical.

The combined light collecting area of a FF sensor might (roughly) be the same, regardless of if there are 20M bigger or 50M smaller pixels.

But how about the combined read noise of 50M pixels compared to the combined read noise of 20M pixels. One might thus think that the sensor having only 40% pixel count would have much less combined read noise? (Or shouldn't we sum read noise?).

The engineers may chime in why - but what I see is that smaller pixels have smaller read noise. In the end, read noise per unit area seems weakly dependent on pixel size.

I had a huge great argument here with Eric on this topic some years ago. Anyhow, this is what I was saying, reworded a bit, in deference to Eric.

When you're looking for dependencies, you end up making some assumptions, which are likely not realistic, but if the reality carries some correlation with the assumptions, maybe you've captured 'weak dependency'.

So, let's assume that all pixels are designed the same (of course, they aren't) and we produce small pixels just by scaling large pixels down proportionately. This means that every feature of the pixel gets smaller (and also, another discussion, that it's fill factor stays the same). The input referred read noise depends on the electronic noise from the pixel source follower and downstream and the conversion gain of the pixel. In turn the conversion gain is inversely proportional to the capacitance of the SF gate and floating diffusion. If these are scaled down the capacitance goes down, the CG goes up and the read noise is reduced.

Now, as said before, pixels aren't in reality designed like that, but we can see why there might be a correlation. For instance:

Companies that have multiple fab lines will tend to use the smaller process nodes for smaller pixel sensors, so the features in those pixel designs will be smaller.

If the sensor is being able to cope with a particular saturation exposure (e.g. to allow use with 100 ISO) then at that exposure the smaller pixel collects fewer photoelectrons, and can be designed with a lower saturation capacity. That means smaller capacitance in the SF gate/floating diffusion.

So if I understand you correctly, you are agreeing with J A C S that smaller pixels should exhibit less read noise given certain plausible but perhaps somewhat artificial assumptions. Is that right?

 bodeswell's gear list:bodeswell's gear list
Canon EOS 90D Canon EOS R Canon EOS R5 Canon EF 135mm F2L USM Canon EF 200mm f/2.8L II USM +4 more
Iliah Borg Forum Pro • Posts: 28,665
Re: Are bigger pixels less noisy?
1

bodeswell wrote:

bobn2 wrote:

J A C S wrote:

JahnG wrote:

Iliah Borg wrote:

JahnG wrote:

Camera manufactors like Canon, Sony, (Panasonic) have made ”low noise” cameras with less and bigger ”pixels”. Do the manufactors not know what they are doing? (Some people in DPR say that such cameras are manufactored because the customers expect that sensors with bigger ””pixels” would be less noisy?)

Suppose 1 large pixel collects the same light as four smaller pixels occupying the same space. Is it possible?

Sounds logical.

The combined light collecting area of a FF sensor might (roughly) be the same, regardless of if there are 20M bigger or 50M smaller pixels.

But how about the combined read noise of 50M pixels compared to the combined read noise of 20M pixels. One might thus think that the sensor having only 40% pixel count would have much less combined read noise? (Or shouldn't we sum read noise?).

The engineers may chime in why - but what I see is that smaller pixels have smaller read noise. In the end, read noise per unit area seems weakly dependent on pixel size.

I had a huge great argument here with Eric on this topic some years ago. Anyhow, this is what I was saying, reworded a bit, in deference to Eric.

When you're looking for dependencies, you end up making some assumptions, which are likely not realistic, but if the reality carries some correlation with the assumptions, maybe you've captured 'weak dependency'.

So, let's assume that all pixels are designed the same (of course, they aren't) and we produce small pixels just by scaling large pixels down proportionately. This means that every feature of the pixel gets smaller (and also, another discussion, that it's fill factor stays the same). The input referred read noise depends on the electronic noise from the pixel source follower and downstream and the conversion gain of the pixel. In turn the conversion gain is inversely proportional to the capacitance of the SF gate and floating diffusion. If these are scaled down the capacitance goes down, the CG goes up and the read noise is reduced.

Now, as said before, pixels aren't in reality designed like that, but we can see why there might be a correlation. For instance:

Companies that have multiple fab lines will tend to use the smaller process nodes for smaller pixel sensors, so the features in those pixel designs will be smaller.

If the sensor is being able to cope with a particular saturation exposure (e.g. to allow use with 100 ISO) then at that exposure the smaller pixel collects fewer photoelectrons, and can be designed with a lower saturation capacity. That means smaller capacitance in the SF gate/floating diffusion.

So if I understand you correctly, you are agreeing with J A C S that smaller pixels should exhibit less read noise given certain plausible but perhaps somewhat artificial assumptions. Is that right?

Ignoring that little amount of noise that may be added due to a more dense sensor infrastructure, mostly yes.

-- hide signature --
bobn2
bobn2 Forum Pro • Posts: 69,811
Re: Are bigger pixels less noisy?

bodeswell wrote:

bobn2 wrote:

J A C S wrote:

JahnG wrote:

Iliah Borg wrote:

JahnG wrote:

Camera manufactors like Canon, Sony, (Panasonic) have made ”low noise” cameras with less and bigger ”pixels”. Do the manufactors not know what they are doing? (Some people in DPR say that such cameras are manufactored because the customers expect that sensors with bigger ””pixels” would be less noisy?)

Suppose 1 large pixel collects the same light as four smaller pixels occupying the same space. Is it possible?

Sounds logical.

The combined light collecting area of a FF sensor might (roughly) be the same, regardless of if there are 20M bigger or 50M smaller pixels.

But how about the combined read noise of 50M pixels compared to the combined read noise of 20M pixels. One might thus think that the sensor having only 40% pixel count would have much less combined read noise? (Or shouldn't we sum read noise?).

The engineers may chime in why - but what I see is that smaller pixels have smaller read noise. In the end, read noise per unit area seems weakly dependent on pixel size.

I had a huge great argument here with Eric on this topic some years ago. Anyhow, this is what I was saying, reworded a bit, in deference to Eric.

When you're looking for dependencies, you end up making some assumptions, which are likely not realistic, but if the reality carries some correlation with the assumptions, maybe you've captured 'weak dependency'.

So, let's assume that all pixels are designed the same (of course, they aren't) and we produce small pixels just by scaling large pixels down proportionately. This means that every feature of the pixel gets smaller (and also, another discussion, that it's fill factor stays the same). The input referred read noise depends on the electronic noise from the pixel source follower and downstream and the conversion gain of the pixel. In turn the conversion gain is inversely proportional to the capacitance of the SF gate and floating diffusion. If these are scaled down the capacitance goes down, the CG goes up and the read noise is reduced.

Now, as said before, pixels aren't in reality designed like that, but we can see why there might be a correlation. For instance:

Companies that have multiple fab lines will tend to use the smaller process nodes for smaller pixel sensors, so the features in those pixel designs will be smaller.

If the sensor is being able to cope with a particular saturation exposure (e.g. to allow use with 100 ISO) then at that exposure the smaller pixel collects fewer photoelectrons, and can be designed with a lower saturation capacity. That means smaller capacitance in the SF gate/floating diffusion.

So if I understand you correctly, you are agreeing with J A C S that smaller pixels should exhibit less read noise given certain plausible but perhaps somewhat artificial assumptions. Is that right?

As I read it, JACS was not saying that smaller pixels should exhibit less read noise, he said that he observed that they did and suggested that 'the engineers may chime in why'. I was chiming in why.

As for the assumptions, whilst they are somewhat artificial, they are no more artificial then the assumptions that tend to underly this kind of discussion.

-- hide signature --

Is it always wrong
for one to have the hots for
Comrade Kim Yo Jong?

Eric Fossum
Eric Fossum Senior Member • Posts: 1,517
Re: Are bigger pixels less noisy?
3

bobn2 wrote:

I had a huge great argument here with Eric on this topic some years ago. Anyhow, this is what I was saying, reworded a bit, in deference to Eric.

Thanks for deference, although to be honest I have no recollection. Hope it was a tie.

When you're looking for dependencies, you end up making some assumptions, which are likely not realistic, but if the reality carries some correlation with the assumptions, maybe you've captured 'weak dependency'.

So, let's assume that all pixels are designed the same (of course, they aren't) and we produce small pixels just by scaling large pixels down proportionately. This means that every feature of the pixel gets smaller (and also, another discussion, that it's fill factor stays the same). The input referred read noise depends on the electronic noise from the pixel source follower and downstream and the conversion gain of the pixel. In turn the conversion gain is inversely proportional to the capacitance of the SF gate and floating diffusion. If these are scaled down the capacitance goes down, the CG goes up and the read noise is reduced.

The limiting noise in SFs these days is often 1/f noise.  1/f noise goes like 1/(gate area)^x where x is about 1 and x depends on a lot of factors it seems. So you get improved CG and worse output-referred voltage noise as the SF gate area shrinks.  CG depends on other parasitic capacitances too so CG improvement is not proportional to SF-gate-area shrink so it could go the other way -> smaller SF gates gives more noise.

I understand you are trying to simplify things here but in the simplification you might get the right answer for the wrong reasons.

Some of you might also be wondering where 1/f noise comes from.  That also depends on the transistor design. In the realm of where we are in QIS-land, it is not traps and RTN, it seems to be mobility fluctuation, which is pretty basic, and the noise is pretty much the same for MOSFETs, buried channel MOSFETs, and JFETs as we are finding out. We also found a way around this! One of my better ideas it seems, which was just accepted for a conference, so stay tuned. I will share when I can. Without the new idea, the best devices have about 0.12e- rms read noise.  That, compared to 0.20e- rms, helps a lot with BER when you are doing photon-counting.

Now, as said before, pixels aren't in reality designed like that, but we can see why there might be a correlation. For instance:

Companies that have multiple fab lines will tend to use the smaller process nodes for smaller pixel sensors, so the features in those pixel designs will be smaller.

If the sensor is being able to cope with a particular saturation exposure (e.g. to allow use with 100 ISO) then at that exposure the smaller pixel collects fewer photoelectrons, and can be designed with a lower saturation capacity. That means smaller capacitance in the SF gate/floating diffusion.

 Eric Fossum's gear list:Eric Fossum's gear list
Sony RX100 II Nikon Coolpix P1000 +1 more
JahnG
OP JahnG Veteran Member • Posts: 3,339
Re: Are bigger pixels less noisy?

Eric Fossum wrote:

bobn2 wrote:

I had a huge great argument here with Eric on this topic some years ago. Anyhow, this is what I was saying, reworded a bit, in deference to Eric.

Thanks for deference, although to be honest I have no recollection. Hope it was a tie.

When you're looking for dependencies, you end up making some assumptions, which are likely not realistic, but if the reality carries some correlation with the assumptions, maybe you've captured 'weak dependency'.

So, let's assume that all pixels are designed the same (of course, they aren't) and we produce small pixels just by scaling large pixels down proportionately. This means that every feature of the pixel gets smaller (and also, another discussion, that it's fill factor stays the same). The input referred read noise depends on the electronic noise from the pixel source follower and downstream and the conversion gain of the pixel. In turn the conversion gain is inversely proportional to the capacitance of the SF gate and floating diffusion. If these are scaled down the capacitance goes down, the CG goes up and the read noise is reduced.

The limiting noise in SFs these days is often 1/f noise. 1/f noise goes like 1/(gate area)^x where x is about 1 and x depends on a lot of factors it seems. So you get improved CG and worse output-referred voltage noise as the SF gate area shrinks. CG depends on other parasitic capacitances too so CG improvement is not proportional to SF-gate-area shrink so it could go the other way -> smaller SF gates gives more noise.

I understand you are trying to simplify things here but in the simplification you might get the right answer for the wrong reasons.

Some of you might also be wondering where 1/f noise comes from. That also depends on the transistor design. In the realm of where we are in QIS-land, it is not traps and RTN, it seems to be mobility fluctuation, which is pretty basic, and the noise is pretty much the same for MOSFETs, buried channel MOSFETs, and JFETs as we are finding out. We also found a way around this! One of my better ideas it seems, which was just accepted for a conference, so stay tuned. I will share when I can. Without the new idea, the best devices have about 0.12e- rms read noise. That, compared to 0.20e- rms, helps a lot with BER when you are doing photon-counting.

 JahnG's gear list:JahnG's gear list
Canon PowerShot A650 IS Panasonic ZS100 Pentax K-r Pentax K100D Super Pentax K-S1
bobn2
bobn2 Forum Pro • Posts: 69,811
Re: Are bigger pixels less noisy?

Eric Fossum wrote:

bobn2 wrote:

I had a huge great argument here with Eric on this topic some years ago. Anyhow, this is what I was saying, reworded a bit, in deference to Eric.

Thanks for deference, although to be honest I have no recollection. Hope it was a tie.

When you're looking for dependencies, you end up making some assumptions, which are likely not realistic, but if the reality carries some correlation with the assumptions, maybe you've captured 'weak dependency'.

So, let's assume that all pixels are designed the same (of course, they aren't) and we produce small pixels just by scaling large pixels down proportionately. This means that every feature of the pixel gets smaller (and also, another discussion, that it's fill factor stays the same). The input referred read noise depends on the electronic noise from the pixel source follower and downstream and the conversion gain of the pixel. In turn the conversion gain is inversely proportional to the capacitance of the SF gate and floating diffusion. If these are scaled down the capacitance goes down, the CG goes up and the read noise is reduced.

The limiting noise in SFs these days is often 1/f noise. 1/f noise goes like 1/(gate area)^x where x is about 1 and x depends on a lot of factors it seems. So you get improved CG and worse output-referred voltage noise as the SF gate area shrinks. CG depends on other parasitic capacitances too so CG improvement is not proportional to SF-gate-area shrink so it could go the other way -> smaller SF gates gives more noise.

I understand you are trying to simplify things here but in the simplification you might get the right answer for the wrong reasons.

Yes, I understand it could go the other way, but we are talking 'weak dependencies' here, so the dependency is in the end about the aggregated chances of it going one way rather than another. Plus, in camera applications maybe a fair part of the read noise is downstream noise, not just the SF, so it's not always the case that the limiting noise in the SF is the one that determines the overall result - and the CG acts on all noises.

The general problem with trying to explain observed trends is that any analysis is bound to be simplistic, because the trend is a whole load of data points which are distributes all around the trend line, and any individual point might be go against the trend.

I think JACS observation aligns with mine, and the conversion gain reason seems to me the most likely, but if there are better suggestions, of course I'll defer to greater knowledge than mine.

-- hide signature --

Is it always wrong
for one to have the hots for
Comrade Kim Yo Jong?

(unknown member) Forum Member • Posts: 61
Re: Quantum efficiency and signal to noise ratio

I never said i am proud of this shot.
It is just a test shot that is almost free of noise at iso 8000 while it manages to keep the details of the actual scene.
So simple as that...

Btw.I am proud of some of these shots..
(All of the shots with D700+ Nikon 28-105 AF-D
https://www.flickr.com/photos/191035018@N07/

J A C S
J A C S Forum Pro • Posts: 19,277
Re: Quantum efficiency and signal to noise ratio
4

Nick Zochios wrote:

I never said i am proud of this shot.
It is just a test shot that is almost free of noise at iso 8000 while it manages to keep the details of the actual scene.
So simple as that...

Actually, there is no much detail preserved. Some of the letters look cartoonish. The texture of the paper on the right is completely gone. The whole scene has no much detail (in focus) in the first place, and there is nothing to compare with, like a non NR version of the shot; or a higher exposure one.

Keyboard shortcuts:
FForum MMy threads