alundeb wroteSay at ISO 100, at average 10000 photons are converted to electrical charge in each pixel. The photon shot noise is then sqrt(10000) = 100 photons, and the signal/noise ratio is 10000/100 = 100If the FWC is doubled to 20,000, then the sqrt = 141.4. Then 20,000 / 141.4 = 141.4 ... for the FWC aspect, correct?
Yep. You obviously can eliminate the division step.
takowasa wrote:
So, for the same exposure and ISO, all sensors will have filled up to the same percentage of their capacity, but sensors with larger capacities will absorb more light in total and thus have less noise.
So ... for a given QE @ constant (same ISO) ... larger pixels will have a larger FWC.
It is beginning to sound as though QE is akin (crude analogy) to using different ND filters, in that the amount of available light is NOT the point of constraint, but rather the amount light allowed to be passed (or efficiency of the conversion to an electrical signal) by different density of absorption (i.e. loss).
So, if the energy contained in the photons in the conversion is NOT 100% efficient, what happens to the inefficiency ... heat (energy conservation @ conversion neither created, nor destroyed).
Until we had a 100% efficient conversion ... THEN the amount of light would be equal to the point of constraint. Currently it is a "QE factored" point of constraint based on the efficiency of the conversion.
takowasa wrote:
So, for the same exposure and ISO, all sensors will have filled up to the same percentage of their capacity, but sensors with larger capacities will absorb more light in total and thus have less noise.
RustyBug wrote:
So ... for a given QE @ constant ... larger pixels will have have a larger FWC.
Not necessarily. For example, the D750 pixels are half the size as the D3s pixels and have almost the exact same FWC (newer tech in the D750 sensor). On the other hand, the FWC of the D600 and D800 pixels (same brand and generation sensors) are in almost exact proportion to their areas.
It is beginning to sound as though QE is akin to using different ND filters, in that the amount of available light is NOT the point of constraint, but rather the amount light allowed to be passed (or efficiency of the conversion to an electrical signal).
Yes. Putting a one stop ND filter over a lens would have the exact same effect as halving the quantum efficiency of the sensor.
RustyBug wrote: takowasa wroteSo, for the same exposure and ISO, all sensors will have filled up to the same percentage of their capacity, but sensors with larger capacities will absorb more light in total and thus have less noise.So ... for a given QE @ constant ... larger pixels will have have a larger FWC.
It is beginning to sound as though QE is akin to using different ND filters, in that the amount of available light is NOT the point of constraint, but rather the amount light allowed to be passed (or efficiency of the conversion to an electrical signal).
All else being equal between two sensors with different pixel counts, larger pixels will have a larger per-pixel FWC but the aggregate FWC will be the same because aggregate FWC = pixel count x per-pixel FWC - this is why pixel count alone doesn't affect shot noise. The constraint is total light along with QE; in good light the extra QE alone may not be needed/useful (although FWC+QE means a lower base ISO and thus better SNR since you can drop the ISO further). In poor light the extra QE will mean less noise.
takowasa wrote:
Not necessarily. For example, the D750 pixels are half the size as the D3s pixels and have almost the exact same FWC (newer tech in the D750 sensor).
It would probably be easier to follow if you wouldn't introduce additional variables somewhat ad hoc ... ie. newer tech in the D750. That's what creates this circular discussion.
I'm assuming that by stating the "newer tech in the D750 sensor", that likely creates a different QE ... which my question was specific to hold the QE constant with only the pixels size variable changed.
I can process and solve for multi-variable algebraic's ... but that really involves reducing to a single variable via substitution before moving on to determine the other variable. Applying differential calculus really isn't my thing as a method of learning for me.
snapsy wrote:
All else being equal between two sensors with different pixel counts, larger pixels will have a larger per-pixel FWC but the aggregate FWC will be the same because aggregate FWC = pixel count x per-pixel FWC - this is why pixel count alone doesn't affect shot noise.
Okay, I'm with you QE / projected light. I'm with you @ calculation for aggregate FWC.
But ... why is it predicated upon the aggregate FWC rather than the pixel FWC?
takowasa wrote:
Not necessarily. For example, the D750 pixels are half the size as the D3s pixels and have almost the exact same FWC (newer tech in the D750 sensor).
RustyBug wrote:
It would probably be easier to follow if you wouldn't introduce additional variables somewhat ad hoc ... ie. newer tech in the D750. That's what creates this circular discussion.
It's not like they build sensors with various size pixels using the exact same tech that I can point to for comparison -- I can only give examples with what exists and has been measured. So, I'm not "introducing additional variables ad hoc" but telling you how things are based on what exists.
So, exactly what is it that you want to know? All I was doing is explaining the relationship between quantum efficiency, the capacity of the sensor to absorb light, and nominal base ISO -- that is, why some cameras have a base ISO of 200, others 100, and others 64 (the D810).
RustyBug wrote: snapsy wroteAll else being equal between two sensors with different pixel counts, larger pixels will have a larger per-pixel FWC but the aggregate FWC will be the same because aggregate FWC = pixel count x per-pixel FWC - this is why pixel count alone doesn't affect shot noise.Okay, I'm with you QE / projected light. I'm with you @ calculation for aggregate FWC.
But ... why is it predicated upon the aggregate FWC rather than the pixel FWC?
Because images are comprised of collections of pixels rather than an individual pixel, so per-pixel statistics don't have much relevance outside of gauging what a full-sized print looks like, which rarely happens. Nearly all images are downsampled to some extent when rendered, either for viewing/printing size or to conform to an aspect ratio, thus it's the per-area stats that matter, esp when comparing noise between two sensors with different pixel densities. An exception to this rule would be when cropping then rendering to a native resolution from that crop (or even upsampling).
The question is a theoretical one to develop the conceptual understanding @ root fundamental math. The fact that there exists a multitude of variability of design in the real world only serves to muddy things a bit ... perpetually introducing changing variables.
I get that the existence of things may not always align perfectly with the question, but the theory still should be able to support it.
RustyBug wrote:
So, if we had a sensor (same physical size) of 1 pixel and a sensor of 4 pixels.
We would have 10,000 photons (100) for the one pixel sensor and 2,500 photons (50) for each of the 4 pixel sensor.
Take me from here to how the 50 @ 4 = 100 @ 1 (ceteris paribus).
When you render the 4-pixel sensor to match the output resolution of the 1-pixel sensor then you'll be downsampling those 4 pixels into 1, which will serve to cut off the higher frequency detail of the 4 pixels (vs the 1-pixel sensor) but also the higher frequency noise (averaging the noise). Someone way smarter than me ( Emil Martinec) did a nice treatment on this subject here. One of Emil's key statements is:
"The linear slope is directly related to the fact that uncorrelated noise of std dev 40 at size 1024x1024 downsamples to std dev 20 at size 512x512, which is the same as the statement that the noise power at 1/2 Nyquist is 1/2 the noise power at Nyquist for uncorrelated noise."
RustyBug wrote:
So, if we had a sensor (same physical size) of 1 pixel and a sensor of 4 pixels.
We would have 10,000 photons (100) for the one pixel sensor and 2,500 photons (50) for each of the 4 pixel sensor.
Take me from here to how the 50 @ 4 = 100 @ 1 (ceteris paribus).
OK, we have a one pixel sensor that has an FWC of 10,000 electrons and a 4 pixel sensor where each pixel has a FWC of 2,500 electrons. Both sensors have a quantum efficiency of 50%. The electronic noise is 4 electrons for the one pixel sensor and 2 electrons for each pixel of the four pixel sensor.
Let's say 3200 photons fall on each sensor, and, for the sake of simplicity, let's assume that this light is uniform. The one pixel sensor releases 1600 electrons. Each of the pixels from the four pixel sensor release 400 electrons.
The photon noise for the one pixel sensor (which would necessarily have to be measured over multiple exposures, but that's a technicality*) is sqrt(1600) = 40 electrons. The combined photon and electronic noise is sqrt(40^2 + 4^2) = 40.2 electrons, since noise adds in quadrature. This results in an image noise of 40.2 / 1600 = 2.5%.
The photon noise for each pixel of the four pixel sensor is sqrt(400) = 20 electrons. The combined photon and electronic noise for each pixel is sqrt(20^2 + 2^2) = 20.1 electrons, which results in a per-pixel noise of 20.1 / 400 = 5%. The image noise, however, is the quadrature sum of the pixel noises: sqrt(4 x 20.1^2) = 40.2 electrons resulting in an image noise of 40.2 / 1600 = 2.5% -- the same as the one pixel photo.
If, on the other hand, the per-pixel electronic noise were the same, the 4 pixel sensor would have slightly higher noise. In the example above, it wouldn't have made a difference. But as the light gets lower, it would have made more and more of a difference, so the sensor with more pixels would produce a photo that was more noisy than the sensor with fewer pixels.
*For a single pixel, we can't see noise at all as there is no variation. We would only be able to see the noise in a single pixel if we took multiple exposures and compared them. For photos made of more pixels, we compute the noise by comparing neighboring pixels where the lighting is uniform and look for the variations in what should be a uniform patch. To complicate things a bit more is the fact that we are talking about Bayer sensors where each pixel records a different color, so, in practice, we need to compare over a much larger patch than 4 pixels. However, these considerations do not invalidate the principles being demonstrated in the thought experiment above.
I would call the photo made with larger pixels "more blurry" rather than call the photo made with smaller pixels "more noisy" when viewed at 100%. And, by the way, note the effect of noise filtering -- it produces a much more favorable balance of noise vs detail than does downsampling.
In other words, even if the 5Ds is intrinsically more noisy than the 5D3/6D at high ISOs, with noise filtering, it may still come out ahead, depending on how much more detail the 50 MP sensor is able to capture over the 22 MP / 20 MP sensors.
takowasa wrote:
The photon noise for each pixel of the four pixel sensor is sqrt(400) = 20 electrons. The aggregate pixel noise is sqrt(20^2 + 2^2) = 20.1 electrons.
How did you get the aggregate pixel noise based on the 20 electrons? Isn't the 20 electrons only for ONE of the four pixels? How does that represent the aggregate?
takowasa wrote:
The photon noise for each pixel of the four pixel sensor is sqrt(400) = 20 electrons. The aggregate pixel noise is sqrt(20^2 + 2^2) = 20.1 electrons.
RustyBug wrote:
How did you get the aggregate pixel noise based on the 20 electrons? Isn't the 20 electrons only for ONE of the four pixels? How does that represent the aggregate?
My apologies for the confusion (I'll go back and edit that post one more time). It was the aggregate photon and electronic noise per pixel, not the aggregate noise for the whole photo.
I would call the photo made with larger pixels "more blurry" rather than call the photo made with smaller pixels "more noisy" when viewed at 100%. And, by the way, note the effect of noise filtering -- it produces a much more favorable balance of noise vs detail than does downsampling.
In other words, even if the 5Ds is intrinsically more noisy than the 5D3/6D at high ISOs, with noise filtering, it may still come out ahead, depending on how much more detail the 50 MP sensor is able to capture over the 22 MP / 20 MP sensors....Show more →
That is correct. A higher MP sensor will have more detail and more pixel-level noise vs a lower MP sensor of the same technology, so you only stand to gain from the higher MP sensor in terms of total IQ, with the possible exception of lower high ISO DR (more noise in shadows due to higher aggregate read noise from the extra pixels, which can't be fully averaged out vs a lower MP sensor due to the asymmetric benefits of analog gain noise reduction (amplifies signal per pixel by 'x' but read noise penalty only occurs once per pixel).
takowasa wrote:
so the sensor with more pixels would produce a photo that was more noisy than the sensor with fewer pixels.
Best to quote the whole paragraph, because a lot of context is lost with that snippet of what I said:
If, on the other hand, the per-pixel electronic noise were the same, the 4 pixel sensor would have slightly higher noise. In the example above, it wouldn't have made a difference. But as the light gets lower, it would have made more and more of a difference, so the sensor with more pixels would produce a photo that was more noisy than the sensor with fewer pixels.
In other words, as I've said all along, more smaller pixels will be more noisy than fewer larger pixels, but this will only be a factor for portions of the photo made with very little light, such as deep shadows or high ISO photography, since the electronic noise is otherwise very small compared to the photon noise. On the other hand, more pixels resolve more detail, giving noise filtering more to work with, so, depending on the particulars of the photo, more pixels with noise filtering may result in less noise for a given level of detail, even for dark shadows and high ISO.