We are currently at a native ISO of 51200 thanks to the 1D X. ISO settings up to 204800, regardless of brand, are "artificial", in the sense that they are only a digital boost above and beyond the maximum native setting.No. The actual native max ISO gain in a 1Dx is below 10,000. Anything and everything above that is digital amplification. This is still pretty impressive, since that is a roughly 100:1 ratio of gain between ISO100 and ISO10,000 - adjustable gains tend to have more inherent noise the larger the range they are supposed to cover is. Canon does NOT indicate "real" ISO's by using real numbers in stead of "Hi" and "Lo" - they indicate the cameras intended usage range. The number is just a number anyway, something that shows in a EXIF or an LCD window.
Actually limiting analog gain to ISO10,000 is a good thing - since taking the analog amplification any higher would actually hurt image performance - increase noise! - so no shame there. I'm actually quite impressed by the noise reduction they apply (but less impressed by the fact that they deny it in official blurbs). As far as I can model it it's a kind of bandwidth energy preserving color NR that works kind of like the Foveon NR system. It is however easily traceable by the noise spread from one channel to another. I have no problem with if they implement it in the "lower cameras" like the 7D replacement.
Last I heard, Canon used two analog amplifications...one at the pixel, which topped out at ISO 1600, one "downstream", but before the ADC (or maybe as part of the ADC itself, never saw a patent), and then, for "expanded" ISO settings, it used digital amplification. In the case of the 1D X, I do not know where the per-pixel amplifiers cut off...it may still be ISO 1600, however I guess I find that unlikely given how good ISO 3200 and 6400 look. If we assume ISO 12800 is the highest the per-pixel amplifiers are capable of, ISO 25600 and 51200 would use the downstream amplifier, leaving H1 and H2 as the only digitally amplified settings.
That jives with past Canon sensors. The 7D is great up to 1600, so-so up to 6400, and terrible at 12800.
There is a finite limit to how far we can really push ISO, simply because the amount of light that will actually reach a sensor in a given very small, finite amount of time is limited.Yes - ISO1,000,000 would mean about 18 electrons per pixel in a 1Dx (still with a bayer filter) with a perfect "100%" QE, at mid gray in the image. That's a S/N ratio of about 1:4, or -2ev - quite useless. One can usually say that the Bayer interpolation of detail gives up at about -2Ev, with most algorithms. But given true hardware-level binning.... You can get reasonable 0.5MP images out of the 135 format sensors at ISO1,000,000. With a separation-style or good quality multi-layer solution you might get a usable 1MP image.
Sure, you could probably "bin", and improve the quality. I guess if you wanted both an excellent low-ISO camera AND the worlds best low-light camera all in one, binning down to half a megapixel would be great. But binning increases the complexity if the sensor, and probably doesn't do as good a job as lower resolution sensor with gigantic pixels.
If someone is really interested in low-light photography...such as indoor events, night sky photography, street photography at night, etc. I think the best option is a real low resolution sensor that just strait up packs the photons into every pixel.
ISO is not really a boost to sensitivity, it is simply a reduction in the white point...it instructs the sensor to register a lesser amount of charge in each pixel as "maximum saturation", or the purest color for each pixel.Well, about. It is an instruction to the ISO amplifier, that sits as an "adaptation gain" between the sensor (pixel) Ve output and the AD-converter's V/ADU input. The sensor in itself is a fixed point, the only thing you can adjust there is pre-load voltages - something typically not done in commercial sensors.
Heh, sure. I was trying to keep lower-lower-level electronics talk out of it, as not everyone here understands things at that level. I might as well put up an electronic diagram and tell everyone to have fun reading it.
Trying to dumb down gain, voltages, etc. just means I end up writing ten times as many words to explain the same thing in a way the majority of people can understand.
For all intents and purposes, ISO settings above 100 simply instructs the camera to read out the sensor with a lower white point. Simplistic, maybe, but all voltages and gain and ADU's aside, that's effectively what's happening.
Assuming an equal exposure value and sensors with equal quantum efficiency (i.e. equal fab tech at the same pixel size), a photo taken at ISO 12800 is going to have four times as much light on the sensor as a photo taken at ISO 51200. Because there is four times less light, at the same Q.E., the ISO 51200 photo will look approximately four times as noisy.No. Increasing absolute exposure by 4x will decrease noise at any given point sufficiently above the electronic noise floor by 2x - not by 4x.
What happens is - taking 18% gray as a point example - that:
18% gray at some ISO has lets say a 400e- signal.
This gives a Poisson of sqrt(400) = 20 = 1:20 = 4.3Ev SNR.
Quadrupling (ISO/4) the base signal gives:
4*400e- = 1600e- of signal.
Pd of 1600e- is sqrt(1600) = 40 = 1:40 = 5.3Ev SNR
-exactly one stop less noise than the four times weaker signal, also at 18% gray.
What happens BELOW that is another matter - you've also increased your DR between 18% gray and the limiting electronic noise floor by 1Ev. If you could get to 3Ev below 18% gray with good detail at the first signal level, quadrupling the signal will give you 4 usable Ev's below 18% gray (if the AD converter and ISO amplifier noise is low enough - this is what's missing in the Canon implementations today, why they have 1.5-2Ev less DR at base ISO).
You are actually quite correct here. I wrote "four times" far too many times in that...blog...
And yes, that is indeed Canon's problem in a nutshell....amplifier noise and ADC noise are really what kill their DR at low ISO. Based on what I've read about Exmor, the use of analog CDS circuits (and non-uniform response inherent to such a design) is also a prime source of vertical banding noise, so it is really the amplifiers, CDS, and ADC that contribute to Canon's low-ISO noise.
Using the 1D X as a basis, with a Q.E. of 47% it has a maximum saturation of 170e- (electrons) at ISO 51200. Assuming 94% Q.E., that would be a max. sat. of 340e-, and for 100% Q.E. a max. sat. of 362e-. Since 100% Q.E. means you are efficiently converting every single photon into an electron, that means 362 photons are captured per pixel. At ISO 1638400, you would capture a mere 11.3 photons per pixel before it saturated!Here, you're missing out on a very basic fact about sensor properties, and how they're reported.
The QE of a sensor is reported AS IS in the construction it is applied.
This means that the 1Dx has got a QE of 47% - this means it converts 47% of the energy already filtered by the CFA. CFA filter average area efficiency is typically around 35-40%. Real incident photon energy > e- conversion rate is then about 0.35*0.47 = 0.16 = 16%... A sensor construction without color filtration would have three times the amount of photons available for conversion.
Hmm, I thought the Q.E. reported by the likes of DXO was "as measured", not "as is". I generally use Sensorgen.info for basic statistics like that, which I believe is based on DXO test results. I'm curious how DXO derives their Q.E., if it is, as you say, what the electronics and photodiode itself (below all the filtration and microlenses and whatnot) is capable of preserving. If it is indeed true that the Q.E. numbers reported on Sensorgen.info (which would really be as reported by DXO) are POST-filtration, then we really do indeed have a LONG way to go before truly preserving ~100% of the light that reaches the sensor through the lens.
To that end, I believe significantly cooling the sensor can greatly help improve the efficiency of the silicon underneath the CFA. I doubt we'll get to -80°C in anything we could buy over the counter, but even moderate cooling to keep the sensor below ambient could improve results, and increase Q.E. as you've described it. There was a rumor that Canon was prototyping such technology as well, which is why I have my hopes up about one of the forthcoming sensor designs from Canon...be it the 7D II or big MP, I think we could see some nice improvements in Q.E. from Canon.
As I mentioned - the Bayer interpolation actually takes away about 2Ev of usable image detail in the end result when all factors are included. A sensor not dependent on filtration - that absorbs all the light you give it - can have more than five times higher e- per photometric exposure level than what the 1Dx has (assuming 80% QE with no color filtration, 60% within the visible band is not unusual in scientific sensors at room temprature)
Aye, completely agree here. Bayer sensors literally throw away unconscionable volumes of light! The problem is, despite the losses from the CFA and interpolation, they still produce better results than the alternatives as of yet. I look forward to seeing more unconventional designs on the market in either Nikon or Canon cameras...something like the Pentax color splitter design in a Canon DSLR would make my day.
The Panasonic patent (that just like all Sony Exmor patents is an implementation patent, not a base patent...) still suffers from some interpolation losses in the color interpolation necessary in the post-processing, even though it uses all available light. AND the usability is severely limited by the angle losses, effective T-stop of a F1.4 lens would be about about T2.5, even if the lens in itself was a perfect T1.4 - so the real gains are only there for F2.8 and smaller aperture lenses. Which kind of takes away a lot of the large-sensor usability. So the 100% light usage sensor is still very far away as you say...
True, there is still interpolation, however I am not sure it suffers "the same" losses as a classic bayer design. At least as I understand the Panasonic patent, there are really only two pixel colors, which would produce a much more even grid of W-R and W+R pixels. The grid layout for either one of those has the same spatial resolution as the green pixels of a bayer array, which has twice the spatial resolution of red and blue pixels. I'd take a more evenly distributed W-R/W+R grid than a bayer array any day. Throw in some efficient cooling to reduce dark current to nearly nil, and I think such a design would perform considerably better than a bayer design.
Regarding the "angle losses", I'm not really sure what you are referring to there...so I'm not sure how to respond.