Is the E-M5 sensor response nonlinear?

[Detail Man]

Thanks for the explanation. I knew most of these things, but when I asked for why raw levels of high ISO (higher amplification, shorter exposure time) images can be higher the missing link was that shot noise indeed does have a temporal component, which in turn is amplified/multiplied by the following analog gain stage. It's the fact that within half the time you can get more than half the electrons that wasn't clear to me when I asked.

As you can see from DxOMark's data, the SNR of photo-cells is on the order of only around 40 dB.

How static has that limit been for the last couple of years and is there a realistic chance to increase this range anytime soon?

(I think) that range of maximum SNR will only increase with higher full-well-capacties and/or higher Quantum Efficiencies. bobn2 is the "guru" on that - perhaps he will see this post and comment.

The MOSFET amplifiers that interface to the photo-cells with each MOS image-sensor photo-site have their own noise (which is on the order of around -80 dB below the voltage signal level that they can roughly linearly process). They also add more (summed) capacitance to the photo-cells that they interface to. bobn2 tells us that it is that (input) capacitance of the MOSFET amplifiers which typically "dominates" over the capactiance of the photo-cells themselves.

Interesting. Within what border does amplification of that 40 dB window happen within the 80 dB range, does it happen close to the clipping point (away from the noise) or rather somewhat below?

Not really sure quite what you are asking. The SNR of the MOSFET amplifier technology (not likely to change a great deal) is around 80 db - which is vastly higher (the linear square ) of the SNR of the photo-cells at there very best (with full illumination due to the highest exposure, lowest F-Number and lowest Shutter Speed).

As ISO Gain increases, the Exposure has to decrease (in order to retain the same "image-brightness"). As a result the SNR of the photo-cells decreasing in those cases (of higher ISO Gains), the 80 dB SNR of the MOSFET amplifiers is always vastly higher than that of the optical photo-cell transducers themselves.

The degradation of the output SNR of the system is (essentially, only) a direct result of the decreased SNR of the photo-cells in lower light levels (and the lower exposure that results from lower light-levels). The only way to improve upon that is by having lenses with lower T-numbers, or the ability to use longer Shutter Times (necessitating tripod-stabilization and low levels of image-sensor dark-noise and thermal-noise sources). With less photons for the image-sensor, those are the only options left to increase Exposure (which does not in itself include the ISO Gain at all).

It's not unlike when trying to use a high-quality microphone and pre-amplifier to record very faint low-level sounds. The SNR of the recording (even with the best equipment) will suffer more and more the lower the sound-pressure-level (SPL) to be recorded.

It's just the square-root function. The Photon Shot Noise varies by that relationship as the amount of light (or the intensity of light per unit area, or the total amount of light transduced by an image-sensor), varies. The square-root of 2 is 1.414, square-root of 4 is 2, square-root of 16 is 4, etc. ...

Thanks for pointing that out again. As you might have guessed I don't care too much for the math, because I usually never really need it. But of course this makes me less of a profound part of an discussion based on these math and the accompanying understanding. There's just not enough room in my head for everything (including remembering things I already learned), so I take the practical route and might have to ask from time to time. Also helps to keep the mind open, but I recognize the additional effort for people I ask these things about. Sorry for that.

 

[Anders W]

It probably ain't that simple, but on the other hand, isn't it almost impossible to get such an almost perfect match by pure coincidence?

The max values are off compared to several camera samples I profiled. So I do not see a match. Noise analysis is absent, RawDigger mode is not stated...

Don't think it's supposed to be the sensors max values

"Green channel (average of G and G2)
3615 1.27 (about 40k pixels clipped)"

Reading, at 3615 saturation was reached, and sensor maxed out.

As usual, you got it wrong

ROTFLMAO

You do not know what you are talking about.

We seem to be rather different creatures Iliah. Unlike you, I do not particularly enjoy rolling on the floor. And unlike you, I do know what I am talking about.

Now where's that evidence in support of E-M5 sensor nonlinearity that you claimed you had? I have lost count of the number of times I have asked you this question. Nevertheless, you just keep dodging without contributing a single bit of interesting information.

Do you ever actually test anything yourself? Or are you just pretending that you do?

 

[Timur Born]

Interesting. Within what border does amplification of that 40 dB window happen within the 80 dB range, does it happen close to the clipping point (away from the noise) or rather somewhat below?

Not really sure quite what you are asking. The SNR of the MOSFET amplifier technology (not likely to change a great deal) is around 80 db - which is vastly higher (the linear square ) of the SNR of the photo-cells at there very best (with full illumination due to the highest exposure, lowest F-Number and lowest Shutter Speed).

What I meant is where the 40 dB photo-cells range window is placed within the 80 dB range of the MOSFET amplifiers? With a SNR of 80 dB and a signal range of only 40 dB it can be anywhere ranging from -80 to -40 dB. The analog amplifiers likely won't run close to their -80 dB noise-floor, do they? I can also imagine that they are rather not run close to their (soft) clipping range, but some dB below that.

That is unless the non-linear analog soft clipping nature is wanted, albeit I doubt that these amplifiers offer a nice and graceful slope and that the digital domain is setup to make use of that. But frankly I have no idea how these things work under the hood. Just came into my mind when these things were mentioned here.

 

[Detail Man]

Interesting. Within what border does amplification of that 40 dB window happen within the 80 dB range, does it happen close to the clipping point (away from the noise) or rather somewhat below?

Not really sure quite what you are asking. The SNR of the MOSFET amplifier technology (not likely to change a great deal) is around 80 db - which is vastly higher (the linear square ) of the SNR of the photo-cells at there very best (with full illumination due to the highest exposure, lowest F-Number and lowest Shutter Speed).

What I meant is where the 40 dB photo-cells range window is placed within the 80 dB range of the MOSFET amplifiers?

The noise of the MOSFET amplifiers is around 80 dB (13.29 EV) below maximum levels. The noise of the photo-cells is (in the brightest illumination possible) around 40 dB (6.64 EV) below maximum levels. The photo-cell noise DOMINATES.

In lower light situations (whether because of illumination-levels, F-Number, or Shutter Time), the Signal/Noise RATIO (as expressed in EV, or in digital "bits") decreases by one-half of the amount of the decrease in the total amount of light that is collected by the image-sensor.

It is (also) necessary to amplify the output of the photo-cells in order to maintain an equal "brightness". For every 6 dB (1 EV) increase in the ISO Gain factor, the noise (including Read/Dark Noise as well as the Photon Shot Noise) that is present is amplified more and more. That is all ...

With a SNR of 80 dB and a signal range of only 40 dB it can be anywhere ranging from -80 to -40 dB. The analog amplifiers likely won't run close to their -80 dB noise-floor, do they?

Please read what I say above.

I can also imagine that they are rather not run close to their (soft) clipping range, but some dB below that.

We are talking about NOISE, not maximum SIGNAL.

That is unless the non-linear analog soft clipping nature is wanted, albeit I doubt that these amplifiers offer a nice and graceful slope and that the digital domain is setup to make use of that. But frankly I have no idea how these things work under the hood. Just came into my mind when these things were mentioned here.

That is about maximum SIGNAL. Linear is what is desired. It is not like a "tube power amplifier".

 

[Great Bustard]

As the amount of light falling on the sensor decreases, the read noise becomes more of a player. So, when we lift the shadows of a base ISO photo, we are seeing more and more of the read noise, and photos taken at high ISOs are mostly shadow that is amplified by the higher ISO, which is why read noise becomes dominant in high ISO photography.

Yes, but the discussion in another thread ( http://forums.dpreview.com/forums/readflat.asp?forum=1041&message=41988325 ) seems to suggest that low ISO read-noise is different from high ISO read-noise (non-ISOless) in a way that is practically relevant. Comparing real images should be the way to get a better "picture" on that instead of just tossing around rather abstract math.

In what way does the linked thread "suggest that low ISO read-noise is different from high ISO read-noise (non-ISOless) in a way that is practically relevant"? From what I gather, it is about if a non-ISOless sensor is using analog or digital gain for intermediate ISO settings.

So, at base ISO, the photon noise is dominant over most of the photo except the extreme shadows. As the light goes down and we raise the ISO to preserve the brightness for any given aperture and shutter speed, more and more of the photo moves to shadow, and the effects of read noise become more and more dominant over photon noise.

While I understand the relation you are describing I am missing one thing: At lower ISO photon noise may be more relevant than read noise...

For the bulk of the photo. But if you are doing a lot of shadow lifting (intentionally exposing lower to preserve highlights), then read noise becomes very significant.

...but with higher ISO/less exposure time you not only get more read-noise affected shadows, but also stronger multiplied photon noise (1x110 = 110, but 2x56 = 112 and 4x 28 = 116). So again it would be interesting to know where all these things come together in practice (images).

The read noise and photon noise contribute equally to a photo when the number of photons hitting a pixel is equal to the square of the read noise divided by the QE of the sensor.

Let's use the GH2 at ISO 3200 as an example:

http://www.sensorgen.info/PanasonicDMC_GH2.html

The read noise is 4 electrons and the QE is 43%. So, when the number of photons hitting the pixel is 4² / 0.43 = 37 photons, the read noise and photon noise will contribute equally.

Another way to think about it, leaving QE out of the equation, is to say that when 4² = 16 electrons are released from the pixel, then the photon and read noise will contribute equally. Since the pixel saturation at ISO 3200 is 598 electrons, this is five stops down from full saturation of the photo.

So, for the GH2 at ISO 3200, the portions five and more stops down from full saturation will be dominated by read noise.

Hope this helps, even if it's not exactly what you were asking about.

Thanks a lot for the good explanation!

I'm glad that you and texinwien found my explanation useful. Apologies for not responding to many of your points, but, as you noted in your post, the answers to those points would require direct measurements that I do not have.

 

[Great Bustard]

Hope this helps, even if it's not exactly what you were asking about.

It does - Thanks!

Thanks for the kind words! If people think of the camera's ISO setting simply in terms of how it affects photon noise and read noise, and what role these two forms of noise play in the visual properties of the final photo, it makes things a lot clearer.

If I may take my explanation a step further, it's important, in terms of comparing systems, to not compare pixel by pixel, but area by area, where the area represents the same proportion of the final photo.

For example, if we are comparing a 10 MP sensor to a 40 MP sensor, we would not compare pixel-by-pixel, but one pixel from the 10 MP camera to four pixels from the 40 MP camera.

If the sensors are the same size and have the same QE, the total amount of light gathered by four pixels of the 40 MP sensor will be the same as one pixel from the 10 MP sensor, thus the photon noise will be the same.

However, the read noise does not add linearly, but in quadrature. Let's say that the pixels of the 10 MP sensor have a read noise of 8 electrons and the pixels of the 40 MP sensor have a read noise of 4 electrons (smaller pixels will tend to have less read noise for a given sensor tech -- strict scaling would result in a read noise of 2 electrons for the 40 MP sensor, but let's say, for this example, that strict scaling cannot be achieved).

Noise, being a standard deviation, does not add linearly, but in quadrature. Thus, the total read noise for four pixels with a read noise of 4 electrons is sqrt (4² + 4² + 4² + 4²) = 8 electrons, which, in this case, is the same read noise as one pixel from the 10 MP sensor.

So, while I kept my explanation at the pixel level in my post above (and below) to illustrate the photon and read noise, in a comparative sense, we have to take it an extra step if the sensors have different pixel counts.

 

[Anders W]

AndersW reported "Ave" statistics.

Exactly. Average values are close without white balance. Field is grossly non-uniform.

Lens is known to the camera, and it is possible the firmware is running compensations.

That's interesting. Some possibility of E-M5 RAW-level "corrections" based on lens characteristics ?

Is that something that you are aware of occuring on other Olympus (or any other brand) camera ?

There is always a possibility that, say, digital vignetting is compensated before raw is recorded. Lots of this stuff happen with some brands, yes, I ran into it. I think it is always good to check for it.

Vignetting correction (except as a by-product of distortion correction and without affect the recorded RAW levels) is known not to happen on any MFT camera unless you turn on the option for "shading compensation". On Oly cameras, that option has an effect on OOC jpegs only. On the GX1 (and possibly some other Panasonic models), it reportedly affects recorded RAW levels too. See here under "vignetting correction":

http://www.dpreview.com/reviews/panasonicdmcgx1/16

 

[Detail Man]

AndersW reported "Ave" statistics.

Exactly. Average values are close without white balance. Field is grossly non-uniform.

Lens is known to the camera, and it is possible the firmware is running compensations.

That's interesting. Some possibility of E-M5 RAW-level "corrections" based on lens characteristics ?

Is that something that you are aware of occuring on other Olympus (or any other brand) camera ?

There is always a possibility that, say, digital vignetting is compensated before raw is recorded. Lots of this stuff happen with some brands, yes, I ran into it. I think it is always good to check for it.

Vignetting correction (except as a by-product of distortion correction and without affect the recorded RAW levels) is known not to happen on any MFT camera unless you turn on the option for "shading compensation". On Oly cameras, that option has an effect on OOC jpegs only. On the GX1 (and possibly some other Panasonic models), it reportedly affects recorded RAW levels too. See here under "vignetting correction":

http://www.dpreview.com/reviews/panasonicdmcgx1/16

It seems that the only way that one could determine whether such a thing takes place would be to shoot without any lens, and compare the result to shooting with a lens attached. Do you know of anyone having tried that in the case of the E-M5? Have you tried that in the case of the E-M5 ?

 

[bobn2]

What I meant is where the 40 dB photo-cells range window is placed within the 80 dB range of the MOSFET amplifiers?

The noise of the MOSFET amplifiers is around 80 dB (13.29 EV) below maximum levels. The noise of the photo-cells is (in the brightest illumination possible) around 40 dB (6.64 EV) below maximum levels. The photo-cell noise DOMINATES.

And in the amplification chain, it is the noise of the first source follower, in the pixel, which is most significant, since that is outside the feedback loop. It can be detected if a camera has two stages of variable gain amplification with two first-stage noise sources, as was done by Emil Martinec here:
http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/noise-p2.html#read_vs_iso

That is unless the non-linear analog soft clipping nature is wanted, albeit I doubt that these amplifiers offer a nice and graceful slope and that the digital domain is setup to make use of that. But frankly I have no idea how these things work under the hood. Just came into my mind when these things were mentioned here.

One can reasonably make the assumption that these companies employ properly trained electronic engineers, who know how to design an amplification chain.

That is about maximum SIGNAL. Linear is what is desired. It is not like a "tube power amplifier".

There are still many who yearn for a return to 'CCD IQ'.

--
Bob

 

[Anders W]

AndersW reported "Ave" statistics.

Exactly. Average values are close without white balance. Field is grossly non-uniform.

Lens is known to the camera, and it is possible the firmware is running compensations.

That's interesting. Some possibility of E-M5 RAW-level "corrections" based on lens characteristics ?

Is that something that you are aware of occuring on other Olympus (or any other brand) camera ?

There is always a possibility that, say, digital vignetting is compensated before raw is recorded. Lots of this stuff happen with some brands, yes, I ran into it. I think it is always good to check for it.

Vignetting correction (except as a by-product of distortion correction and without affect the recorded RAW levels) is known not to happen on any MFT camera unless you turn on the option for "shading compensation". On Oly cameras, that option has an effect on OOC jpegs only. On the GX1 (and possibly some other Panasonic models), it reportedly affects recorded RAW levels too. See here under "vignetting correction":

http://www.dpreview.com/reviews/panasonicdmcgx1/16

It seems that the only way that one could determine whether such a thing takes place would be to shoot without any lens, and compare the result to shooting with a lens attached. Do you know of anyone having tried that in the case of the E-M5? Have you tried that in the case of the E-M5 ?

How could one determine whether vignetting correction takes place by comparing results obtained without lens with results obtained with lens? As far as I can tell, such a comparison would tell us nothing about vignetting correction. Nor can it, as far as I can see, tell us anything else of interest in this context. So no, I didn't try shooting without a lens and wouldn't want to try it unless you can FedEx me a dustfree chamber.

I do have one of my legacy lenses with me up here in the pre-Apennines but I wouldn't want to reshoot the test with that either unless someone can come up with a convincing reason for doing it.

As to the linearity of sensor response, it is not of much interest whether vignetting correction takes place or not. Iliah Borg, as usual, fails to specify a credible mechanism whereby such correction would make a test like mine fail.

If vignetting correction were indeed taking place, that correction would of course be based on the lens used, the aperture used (and the FL used if we are talking about zooms). I of course made sure to keep all of those things constant throughout so how could they influence the way RAW levels increase with exposure (shutter speed)?

 

[Iliah Borg]

It probably ain't that simple, but on the other hand, isn't it almost impossible to get such an almost perfect match by pure coincidence?

The max values are off compared to several camera samples I profiled. So I do not see a match. Noise analysis is absent, RawDigger mode is not stated...

Don't think it's supposed to be the sensors max values

"Green channel (average of G and G2)
3615 1.27 (about 40k pixels clipped)"

Reading, at 3615 saturation was reached, and sensor maxed out.

As usual, you got it wrong

ROTFLMAO

You do not know what you are talking about.

We seem to be rather different creatures Iliah.

Yes, I do not start dialogues with complete distortions of the words of others, I know how to make accurate experiments and how to present data, and I know the difference between linear and linearized. Somebody like you saying "The advantage of the approach chosen by Olympus is that photographers who shoot RAW run less risk of clipping the highlights by mistake." can't deal with any data

--
http://www.libraw.org/
http://www.RawDigger.com/

 

[Iliah Borg]

One can reasonably make the assumption that these companies employ properly trained electronic engineers, who know how to design an amplification chain.

Assuming this one can go as far as realizing that camera meter calibration to ETTR-0,5..-1,5 and highlight clipping in unadjusted conversion like OOC JPEGs is not because all camera makers are idiots

--
http://www.libraw.org/
http://www.RawDigger.com/

 

[bobn2]

One can reasonably make the assumption that these companies employ properly trained electronic engineers, who know how to design an amplification chain.

Assuming this one can go as far as realizing that camera meter calibration to ETTR-0,5..-1,5 and highlight clipping in unadjusted conversion like OOC JPEGs is not because all camera makers are idiots

Are they not? I'm sure I can find many posts from these forums saying that they are, which, on the doctrine that frequently expressed views must be right, proves the case
--
Bob

 

[Timur Born]

Please read what I say above.

...

We are talking about NOISE, not maximum SIGNAL.

Yes, but I was talking about signal. I just was curious where the amplifier might be working within its own limits. When you have a (possible) gain range of 80 dB and only need to amplify a signal range of 40 dB then it makes sense not to use an analog amplifier too close to its own ceiling.

That is about maximum SIGNAL. Linear is what is desired. It is not like a "tube power amplifier".

Surely not a tube power amplifier, but an analog amplifier nonetheless, which usually means that it may behave non-linear towards its upper limit. This is why I was curious how "hot" these amplifiers are run while they have so much room for playing it save.

 

[Iliah Borg]

One can reasonably make the assumption that these companies employ properly trained electronic engineers, who know how to design an amplification chain.

Assuming this one can go as far as realizing that camera meter calibration to ETTR-0,5..-1,5 and highlight clipping in unadjusted conversion like OOC JPEGs is not because all camera makers are idiots

Are they not? I'm sure I can find many posts from these forums saying that they are, which, on the doctrine that frequently expressed views must be right, proves the case

ROFL. At least Adobe are not helping to enforce ETTR.

--
http://www.libraw.org/
http://www.RawDigger.com/

 

[Anders W]

It probably ain't that simple, but on the other hand, isn't it almost impossible to get such an almost perfect match by pure coincidence?

The max values are off compared to several camera samples I profiled. So I do not see a match. Noise analysis is absent, RawDigger mode is not stated...

Don't think it's supposed to be the sensors max values

"Green channel (average of G and G2)
3615 1.27 (about 40k pixels clipped)"

Reading, at 3615 saturation was reached, and sensor maxed out.

As usual, you got it wrong

ROTFLMAO

You do not know what you are talking about.

We seem to be rather different creatures Iliah.

Yes, I do not start dialogues with complete distortions of the words of others,

Neither do I as I have already demonstrated here:

http://forums.dpreview.com/forums/read.asp?forum=1041&message=41996611

But in contrast to you I am used to backing up my claims by evidence.

I know how to make accurate experiments and how to present data,

I haven't seen the slightest sign of that yet. Just confirmation of the old proverb that empty barrels make the most noise.

and I know the difference between linear and linearized.

So do I. So what?

Somebody like you saying "The advantage of the approach chosen by Olympus is that photographers who shoot RAW run less risk of clipping the highlights by mistake." can't deal with any data

What I said is obviously true and so provides yet another demonstration of your lack of proper comprehension. That is the advantage of the approach chosen by Olympus. The disadvantage is that photographers who shoot RAW might also fail to minimize noise. Of course, and as I have already pointed out in the thread you are referring to, both the advantage and the disadvantage can easily be nullified by anyone who cares enough to find out how the meter is calibrated in relation to sensor saturation.

 

[Iliah Borg]

It probably ain't that simple, but on the other hand, isn't it almost impossible to get such an almost perfect match by pure coincidence?

The max values are off compared to several camera samples I profiled. So I do not see a match. Noise analysis is absent, RawDigger mode is not stated...

Don't think it's supposed to be the sensors max values

"Green channel (average of G and G2)
3615 1.27 (about 40k pixels clipped)"

Reading, at 3615 saturation was reached, and sensor maxed out.

As usual, you got it wrong

ROTFLMAO

You do not know what you are talking about.

We seem to be rather different creatures Iliah.

Yes, I do not start dialogues with complete distortions of the words of others,

Neither do I

It does not matter what you say now, you were caught on the spot. It only adds to fun watching you mixing elementary things up. But please continue, it is fun indeed.

Somebody like you saying "The advantage of the approach chosen by Olympus is that photographers who shoot RAW run less risk of clipping the highlights by mistake." can't deal with any data

What I said is obviously true

But of course it is not true

--
http://www.libraw.org/
http://www.RawDigger.com/

 

[Anders W]

It probably ain't that simple, but on the other hand, isn't it almost impossible to get such an almost perfect match by pure coincidence?

The max values are off compared to several camera samples I profiled. So I do not see a match. Noise analysis is absent, RawDigger mode is not stated...

Don't think it's supposed to be the sensors max values

"Green channel (average of G and G2)
3615 1.27 (about 40k pixels clipped)"

Reading, at 3615 saturation was reached, and sensor maxed out.

As usual, you got it wrong

ROTFLMAO

You do not know what you are talking about.

We seem to be rather different creatures Iliah.

Yes, I do not start dialogues with complete distortions of the words of others,

Neither do I

It does not matter what you say now, you were caught on the spot. It only adds to fun watching you mixing elementary things up. But please continue, it is fun indeed.

The laugh is entirely on you. Feel free to continue making a fool of yourself.

Somebody like you saying "The advantage of the approach chosen by Olympus is that photographers who shoot RAW run less risk of clipping the highlights by mistake." can't deal with any data

What I said is obviously true

But of course it is not true

--
http://www.libraw.org/
http://www.RawDigger.com/

 

[bobn2]

Please read what I say above.

...

We are talking about NOISE, not maximum SIGNAL.

Yes, but I was talking about signal. I just was curious where the amplifier might be working within its own limits. When you have a (possible) gain range of 80 dB and only need to amplify a signal range of 40 dB then it makes sense not to use an analog amplifier too close to its own ceiling.

That is about maximum SIGNAL. Linear is what is desired. It is not like a "tube power amplifier".

Surely not a tube power amplifier, but an analog amplifier nonetheless, which usually means that it may behave non-linear towards its upper limit. This is why I was curious how "hot" these amplifiers are run while they have so much room for playing it save.

An amplifier running with 40dB of negative feedback is linear right up until where it clips. That's the difference between 'transistor sound' and 'tube sound'.
--
Bob

 

[Iliah Borg]

Feel free to continue making a fool of yourself.

I knew you can't stop. On a side note, fools are not qualified to make judgements on the foolishness of others. Of course I do not expect you to understand that, but maybe you can at least practice that simple rule?

--
http://www.libraw.org/
http://www.RawDigger.com/