Printable View
I know in the film days there was ASA, which evolved into ISO in digital cameras, and that there is a actual ISO that dictates the standards. But how exactly does Canon (or Nikon for that matter) determine that 51,200 is one stop less than 102,400 and that 204,800 is one stop more? I guess I'm looking for a lay person explanation if it's possible.
F stops represent a 100% change in light. So 1 stop is either 1/2 the light when closing or 100% when opening. Same shutter speed, one stop closed equals 1/2 the light hitting the sensor. One stop more open doubles the light hitting the sensor. So when working the algerbra to ISO, when you want to move the Fstop one down (closed) then you have to double your ISO to get the same exposure. ISO 200 would then have to go to 400, then 800, 1600, etc.
That's basically it. ISO 100 is an approximation to ISO 100 film in sensitivity (the same aperture and shutter speed should give the same exposure, in theory). Of course it's slightly more complicated than that due to sensor noise, marketing "statistics", and the like, but some sites like DPReview will even go out of their way to test the accuracy of the ISO ratings of each camera. I suppose the idea is that if ISO 51,200 film existed, it would have the same sensitivity to light as a digital sensor set to ISO 51,200.
Thanks BK and Positron, okay I understand that. But how does Canon determine that for example the 1D X's sensor sensitivity to light is actually ISO 102,400 and not 204,800, if we were to set it to those settings? I'd say, well because one has more noise in the image. But is that how do Canon measures the ISO differences and determines uniformity between models, etc.
@Bigblue That's a function of what's called the quantum efficiency of the detector. Basically, what's the probability of creating an electron when a photon strikes the detector. Then there's the amount of charge that each pixel can hold before it's saturated (the value won't go up anymore as more photons hit). I'm guessing that the 50% mark is hit at similar exposure conditions that would give 50% gray on equal ISO film. There may be some other fraction of the total amount that's agreed upon by the digital camera manufacturers but the idea is the same. If you could magically create an array that could store 100x more photons then you could declare it to be ISO 1 instead.
Good question - at what point does noise overtake signal to some agreed upon level. WIth Magic Lantern you can push the sensor to stupid effective ISO that creates about 50% noise and 50% signal.
As sensor technology as well as algorithms for interpolating data (in firmware) improve, the ability to squeeze more usable pixels out of the camera at high ISO improves. If Canon wanted, they could release a camera today with ISO ten billion just by taking the image at base sensitivity (assume ISO 100) and ratcheting the exposure up +27 EV (give or take). It would look similar to what would happen if you tried the same thing in Photoshop -- not pretty. They always play a balancing game between pushing the highest numbers possible for marketing purposes and trying to only give the user the option of what's vaguely usable so customers don't get disgruntled when they buy a camera that can go up to ISO 13,421,772,800 and the images look terrible... it's bad for business. :D
@bigblue1ca. Well I asked the same question some years back of Chuck Westfall (Canon) and he supplied me with the Standard Specification "Sensitivity of Digital Cameras CIPA DC-004 (2004)". I would be happy to send it to you (400K PDF) as a PM, but I don't think its possible, any ideas?
Interesting all, thanks. Raid, sure that would be great, you could just add @gmail.com to my username and email to me if that works for you.
Or, you could click HERE for the pdf.Quote:
Originally Posted by bigblue1ca
As for, 'how does Canon know that ISO 102400 is one stop more than ISO 51200', put in lay terms, it's simple math. The sensor detects photons of light, each pixel 'counts' those photons for the selected exposure time, storing them as electrical charge. If the ISO is set to 100 (let's assume that's the 'base' - any sensor has only one base ISO), that amount of charge is converted to a digital signal of a certain numerical value, which is the value assigned to that pixel for brightness (the position under the RGB mask determines the color that's assigned). If the ISO is higher than 100, up to the top 'native' ISO (i.e. non the expanded 'H' or 'L' settings), the sensor applies an analog gain to the signal prior to converting it to digital. Analog gain is like 'turning up the volume' on the signal electrically, and the amplifiers are precise, unlike turning the know on your stereo, it can apply a 2-fold amplification for ISO 200, a 16-fold for ISO 800, etc. If you've selected an expanded ISO like ISO 102400 on the 1D X, that is a digital gain applied after the analog signal has been converted to digital. So for the ISO 102400 case, the camera is applying the maximum analog gain (to get to ISO 51200), then it just takes the numerical value assigned and doubles it to reach ISO 102400.
Thanks Neuro that makes sense.
This would be somewhat problematic since Canon only has a 14 bit raw :)Quote:
Originally Posted by Positron
The quantization of the voltage is independent of the amount of gain added.Quote:
Originally Posted by Jon Ruyle
There is enough leeway in the "standard" that manufacturers can pretty much make up whatever they want for the "ISO" -- and they do. The ISO standard did include a simple and obvious method that everyone should use, but no one does. Too bad, because it would have been nice to be able to go from one camera to the next without having to spend the time figuring out what the iso setting *really* means. Many manufacturers, like Canon, aren't even consistent across model lines in the same generation, let alone different generations and different manufacturers.Quote:
Originally Posted by bigblue1ca
Ideally, sure. But Canon currently achieves ISO above 1600 by pushing the raw.Quote:
Originally Posted by ChadS
Is that true for all cameras? I thought Canon defined 'native' ISOs as analog gain and 'expanded' ISO as digital gain ("pushing the RAW").Quote:
Originally Posted by Jon Ruyle
It was true for all cameras until a few years ago.Quote:
Originally Posted by neuroanatomist
Not really. There is a strong correlation between native ISO and analog gain, but there are several exceptions. For example, on the 5D2, 6400 is a 'native' ISO, but it is based on digital gain (3200 is the highest analog-only gain).Quote:
Originally Posted by neuroanatomist
Personally, it doesn't really matter what the highest analog gain is -- I'm more interested in the highest *useful* analog gain. For example, the D7000 has analog gain up to ISO 1600, but going above ISO 200 in raw has only downsides -- no benefits. (Same with going above 1600 on the 5D2 -- although there is a slight reduction in pattern noise, for me it's not usually worth the loss in highlight headroom.)
I don't know for sure that 1600 is the highest real iso on all of them, but a friend who has examined raw files for more than one canon camera (including the 5DII) tells me that for those he has looked at, everything above 1600 is achieved by pushing the raw. Thus it is definitely not the case that pushing the raw only happens with "expanded" iso settings.Quote:
Originally Posted by neuroanatomist
Really? I was told 1600 was the highest.Quote:
Originally Posted by Daniel Browning
If 3200 is the highest analog-only gain, then what do you mean when you say 6400 is 'native'? Do you simply mean it is not an expanded iso?
Yes, that's what I meant.Quote:
Originally Posted by Jon Ruyle
So...is there any hardware basis for the distinction between native and expanded (H#) ISO settings for Canon bodies, or is it all arbitrary?
As far as I can tell, it's either monkeys on typewriters or marketing VP's with dartboards. :DQuote:
Originally Posted by neuroanatomist
lol. :DQuote:
Originally Posted by daniel browning
Hi Guys....ok....so usually once we get to the point where we are trying to differentiate between a monkey on a typewriter and a marketing VP with a dartboard, the thread ends.....and rightfully so.....but I have found this very interesting and wanted to bump the thread as this has started to make some sense to me....
This thread has reminded me of some comments Daniel made awhile ago about how Canon should decrease file size with increasing ISO as some of the data was becoming useless (feel free to correct me if I am off) and dynamic range was decreasing. This thread is helping me understand what may have been meant by that statement. What I am trying to wrap my head around is how ISO works and the relationship between ISO and dynamic range.
I do work a (very) little with electronic parts and analog signals. Our devices have a typical analog signal from 4 to 20 mA (I assume that the sensors report a different range and, maybe, units). So, is what you are saying that each pixel reports an analog signal within the range (say 4 mA) and when the camera is set at ISO 100, it uses 4 mA to represent the intensity of that pixel in generating the image? Then, for ISO 200 it doubles the 4 mA signal output and uses 8 mA to generate the image, etc? In that event, the dynamic range associated with an output at the top of the range is lost. Correct?
Is this what you mean by adjusting analog gain? If so, it is pretty easy to see how the DR would become limited. I know I am only using 4 to 20 mA because that is what I am familiar, but if the dynamic range is equivalent to the analog output at ISO 100, then wouldn't the top half of the data be lost with every stop gained by adjusting ISO? That is an awful lot, so I suspect the relationship between analog output and DR isn't linear. Is this off-base? If so, could someone elaborate on the relationship between the analog output and DR?
Of course, after I understand that, the next questions are what is stored in a RAW file if it isn't the analog output and what is meant by "pushing RAW?" Quick guess is that we are going from analog scale (similar to the 4-20 mA) which would be limited by the sensitivity of the device to a digital scale (somewhat limitless).....but this is why am asking the question.
Thanks,
Brant
@Kayaker you're close. Photons that strike the pixel create electrons - but only a few. I then have an amplifier circuit that amplifies the number of electrons by some multiplier. That's the gain. I accumulate these electrons in something called a charge well (think capacitor). The more electrons the higher the voltage (as opposed to current that you mentioned).
Now, I've got a voltage and I need to convert that to a digital number. For this we use an A/D -> Analog-to-Digital converter. Basically an A/D simply represents voltages as steps from a minimum to a maximum. If I had a 4 bit (16 different values) A/D ranged from 5V to 6V I would read 0000 at 5V and 1111 at 6V and 1000 at 5.5V, etc.
I have absolutely no idea what the actual range of voltages used in the Canons but I do know my camera has a 14-bit A/D. That is, in theory, there are 16,384 values from the lowest voltage to the highest voltage that it can represent. The last digit in the number represents the "Least Significant Bit" (LSB). In other words, the voltage change that corresponds to changing the output by a count of 1 is the smallest voltage change we can record. However, the amount of noise in the signal is generally of the order of the LSB or higher. In such a case were I to simply truncate all my numbers and treat the last number as 0 there would be no noticeable change in the result.
In low-light conditions none of the sensors are going to have a very large voltage on them. In such a situation the first several digits of every pixel will be zero. I suspect that by pushing the raw they're saying that instead of leading with a bunch of zeros the lower digits are pushed left - and the numbers are padded with 0's at the right end (like representing 4, 5 and 8 as 400, 500, and 800 - all the ratios are the same). Doing so means that a bunch of information is all 0's and doesn't really need to be written. If I know that the last 7 digits of every value are 0 I should just be able to make a note of that and only record the 7 useful digits from each pixel.
However, I suspect that the complexity of changing the file format for the built-in JPEG preview maker and the part of the code that writes to the card would become overly complex or else we'd have seen that.
So let me wander the topic - if we went to a 16 bit D/A we would have greater dynamic range?
Could the origninal ISO question be answered with the minimum light needed to achieve a certain charge value?
@BK Yes, and No.
Dynamic range is (by definition) the ratio of the greatest signal that can be properly recorded to the least signal that can be properly recorded. However, going beyond 12 bits in non-scientific-grade equipment is non-trivial. Most of the time the last couple of bits in your photos aren't valid. You've probably got 11-13 real bits (vs. noise or artifacts) depending on conditions for ISO 100. Any higher amplifier gain and you'll lose SNR - the noise floor increases but the cap does not - hence you lose dynamic range. Scientific-grade camera have active cooling elements called Peltier Coolers on them that keep the amplifiers well below freezing (or sometimes they are even used with a liquid nitrogen supply). Unfortunately, the power the Peltier cooler has to reject plus the power required to run it create a massive heat load that must be blown into the air. Can you imagine a yet-to-be-announced 5diii with a 6" heat sink hanging off to one side of it with a fan? It's not going to happen... However, scientific grade cameras can easily surpass 16 significant bits. I believe there are companies out there peddling 20 or 24-bit cameras. And such a camera is capable of an astounding dynamic range that would blow your SLR away. And they should for $250k - $1M apiece!
There's an old technology called CID (charged-injection device) elemtns vs. today's CCDs or CMOS sensors. In a CID camera you could saturate every pixel and it wouldn't bleed into the pixels around it. Furthermore, you could tell when this saturation happened according to a clock (and turn off the amplification to that pixel as a result). The neat thing about this is that your dynamic range is no longer limited by the bit depth of the sensor but rather on the clock you use. The astronomy guys loved these things. They could saturate the image of the stars in their field of view in a matter of seconds or minutes but leave the exposure on for hours and hours to get the background gasses, etc. I don't know why this technology fell by the wayside except that maybe the mass market for CCDs and CMOS sensors just makes inferior sensors so much cheaper it's not worth pursuing CIDs anymore.
As to ISO, increasing the dynamic range cannot equate to sensitivity. I didn't bother to read that PDF that was linked but ISO relates to saturating your sensor (100% white) or 50% gray or 19% gray at various light levels. Adding bits doesn't happen at the top end. So a 14 bit sensor and a 16 bit sensor might still be set so they max at the same exposure to the same light. The 16 bit sensor would just do a better job of distinguishing 1% gray from 1.01% gray.
Essentially, yes, that's correct.Quote:
Originally Posted by Kayaker72
Your suspicion is correct. First, in Canon cameras, the non-linearity is caused by the very poor ADC performance (compared to any modern non-Canon DSLR). The ADC adds so much noise that the first 6dB of amplification (from ISO 100 to ISO 200) reduces dynamic by less than 1dB -- it's almost exactly the same. That's one reason why many Canon photographers consider ISO 100/200 to be interchangeable. But for successive analog gains, the ADC's contribution to noise continues to shrink. Above 1600, the ratio between the amount of gain and the amount of lost dynamic range becomes 1:1. Second, the input signal (light) is often varied with the gain level, such that the top half of data is not lost. Third, most photographers only use between 5 and 7 stops of dynamic range anyway, so you can throw a lot away before they start to notice.Quote:
Originally Posted by Kayaker72
Well, Canon likes to apply both analog and digital gain to the raw file, depending on settings and lenses.Quote:
Originally Posted by Kayaker72
Yes, it's digital gain as you guessed. It can be either linear (chop off the top half and throw it in the garbage) or nonlinear (compress the top half into fewer values; HTP is an example of this).Quote:
Originally Posted by Kayaker72
BTW, great post, ChadS.
No. Actual instances of a camera being limited by the ADC bit depth are extremely rare. Increasing the bit depth of the ADC is trivially easy. Mass market DSLR have already shipped with 21-bit ADC -- and they could build one with a 1000-bit ADC if they wanted to. Increasing the bit depth alone is easy -- it's like adding a new speedometer to your vehicle. But putting a 500 MPH speedometer on my Pinto isn't going to make it any faster. :) The hard part, of course, is to build a camera that can actually utilize the higher bit depth. Just like it's hard to make my Pinto actually go 500 MPH.Quote:
Originally Posted by Busted Knuckles
Canon's noise problems are so severe, they could actually go to 11-bit files without any loss of usable dynamic range. They jumped to 14 bits before their cameras were even making full use of 12 bits. Ah, the joy of marketing.
If you mean ADU (i.e. the numbers in the raw file), then yes, if it's scaled for bit depth or given as a percentage of the maximum value. But if you meant pre-ADC charge units, like mV, then I don't think it would as useful, because in most situations we'll be more concerned with what actually makes it to the file.Quote:
Originally Posted by Busted Knuckles
Ahhh somewhat clearer, it sounds like bits do not help the absolute range, but would help in the gradations between the absolutes - the question is what is a praticable limit. Back in the day.... the zone system (preferable is formats larger than 35mm) we would have many more zones and EV values within a single negative than what I have been able to produce w/ the electronic sensor - maybe it is just me and I haven't figured it out yet.
Back to ISO capability. The lower amplificaton levels would improve total dynamic range - noise floor stays down? i.e. what happened to the 50 ISO setting? The next question is for a given level of amplification and light, what is the signal - charge level - produced i.e. measure of sensitivity and ultimately S/N = max iso?
Now the question is how to get the sensor to be more sensitive needing less amplification???
I am so confused
@BK I defer to Daniel's expertise on Canon's A/D converters (ADCs). He stated that for lower pre-amp gains, Canon's ADCs are terribly noise and non-linear. So by his estimate ISO 100 and 200 are basically the same. In a perfect world we could have ISO 10 for really sunny days at f/1.2. However, there's a minimum amount of amplification necessary to actually get the signal moving - hence no super-low ISO. Some medium format digital sensors do offer sub-100 ISO values. This makes sense as their sensor areas are larger as are their pixels (generally) so there's more photons to work with at the lowest levels of amplification.
So, what do we do to get more signal from fewer photons? That's the $1M question. You've heard of backside illuminated sensors? That's one way. Improving the gain/noise in the pre-amps is another. Reducing the spurious noise sources helps (active cooling above), better energy-level matching for the semiconductor junctions to photon energy... There's a bunch of technologies at improving CCDs, CMOS, etc. Alas, the really advanced stuff is not destined for SLR for many, many years after it's available in scientific-grade sensors for quite a while. True, the market is huge for pro / pro-sumer / enthusiast DSLRs but the margins are thin. The scientific instruments have huge margins but very thin sales. But no self-respecting researcher is going to pay a 10x markup (or more) for a camera that's no better than what he can buy from his local box store. So the sci-grade stuff has to be that much better.
Now I know its been a very long time since I went to school (chalk on slate).
Dynamic Range (DR) is a ratio between the maximum and minimum values.
Everybody is focusing on minimum value around the noise. You can also increase the DR by increasing the maximum, that is being able to store more charge in each pixel.
Sorry if I'm being difficult
Tony
@Tony that's effectively what you're doing when you change the gain. Let's say the maximum is 1000 units. If I'm receiving 1 unit per 1/1000 and I multiply by 100 or if I'm receiving 10 units per 1/1000 and multiplying by 100 I'm still maxed.
Having no idea what the charge well depth is for any individual sensor we can only make broad generalizations. However, I will say that adding a larger capacitor for each pixel is inviting more dark noise so that's not all positive. Given that, every amplifier is tuned to nearly max the storage ability of whatever the well can hold. So one can't really increase the DR by increasing the max - that's already assumed since amplifying to anything less than the max is sacrificing range. Hence, the interesting bits happen in the noise.
@Chads
http://www.andor.com/learning/digita...ras/?docid=321
This link give Dynamic Range and Full Well Capacity of some commercial CCDs. While the well capacity and noise is provided I'm not sure about the units.
Although I do like all of those dB values, wish Canon would do that, it would make sense to me.
Tony
Unfortunately, the bits don't help that, either. The gradations are limited by noise (photon shot noise) to less than 8 bits, even in the very lowest-noise circumstance (ISO 100, ample light). The only thing that helps gradations that is increased full well capacity (FWC) when shooting in ample light, and sensitivity when shooting in low light.Quote:
Originally Posted by Busted Knuckles
I'm not sure, but the ideal would be 1 bit -- to record the location of every photon strike individually. Then your effective bit depth would vary by output size, from 1 bit up to 16 bits or more.Quote:
Originally Posted by Busted Knuckles
It's not just you. Right now the primary limitation is Canon and typical raw converters. With non-Canon cameras and a good raw converter (e.g. RawTherapee), you can get a full 13 stops dynamic range -- 7 above middle gray and 6 below -- just like the best film (and careful chemical development). But most photographers prefer to stay in the 5-7 stop range because it gives the image "punch" and "pop".Quote:
Originally Posted by Busted Knuckles
It sounds like you're asking if dynamic range would improve by removing the amplification altogether. In general, yes. That's what a CCD is -- the FWC there is limited only by the photodiode itself (so the lowest ISO setting is fixed by the the second place where the energy goes, after knocking an electron loose from the depletion region). In a CMOS sensor, on the other hand, the FWC can be limited by the next link in the chain, one that CCD does not have, the source follower transistor. With a low capacitance, CMOS can suffer very poor FWC and maybe low dynamic range (certainly low max SNR). But designers also have the choice of high capacitance, and the FWC is generally as good as CCD.Quote:
Originally Posted by Busted Knuckles
As for the ISO 50 setting, the fundamental lower limit of the ISO setting is set by the full well capacity and sensitivity. If you increase sensitivity without increasing FWC, your minimum ISO must increase. But you can always put an ND filter on there and get the exact same result as before. But if you increase the effective FWC somehow (such as with an HDR pixel or even just software), then you can get significantly reduced noise or increased dynamic range with lower ISO settings.
I'm not sure what you mean here, but one example might be that on some cameras, 10k electrons gives you the level that corresponds to middle gray in the default raw conversion of some converters. To calculate the S/N, you factor in read noise and photon shot noise, and then the max "usable" ISO could be based on whatever minimum SNR you prefer (I'm happy with 0dB, but a lot of photographers prefer 10dB minimum).Quote:
Originally Posted by Busted Knuckles
There's not really much left to be done. If you don't count color filters, Modern sensors are already capturing 70-80% of the incident light (at the green wavelength at least). Even theoretically, it's only possible to increase to 100% -- since you can't see light that isn't there, of course. The only big jump would be if we found some other way of sensing color (rather than the filters we currently use).Quote:
Originally Posted by Busted Knuckles
Where there is a *lot* of room for improvement is in the read noise -- the part added by the sensor. If there were no read noise, we could shoot at ISO 1,000,000 and still be able to see detail (it would be very, very noisy, but still plainly visible).
The units are "photo-electrons (electrons knocked loose by an incident photon) or their equivalent (e.g. an electron caused by internal thermal energy instead)".Quote:
Originally Posted by Raid
I do too. It would be nice if they also stopped spouting bald-faced lies (e.g. "our 14-bits give improve gradations!"), but I'm not holding my breath.Quote:
Originally Posted by Raid
Bleh - we need to get rid of CCD and CMOS and replace all 20 million pixels with photomultiplier tubes! Single photon counting is not a problem!
Ok, seriously though, I was unaware that the quantum efficiency of CCDs was >70%.
There is another way to sense color other than filters and the solar power folks are doing it now. Put your highest bandgap doping at the top of the stack, and reduce the gap for each progressively lower layer. Commercial solar power chips use 3 layers but there's no limit - so long as you're not absorbing the photons in the silicon crystal. This matches the energy of the photon to the absorption band of the semiconductor. For solar cells the result is more power, less heat. For cameras it would be a huge increase in range and also provide a better approximation of color - actually it could provide a low-res spectrum at each pixel.
@Daniel
"Well Capacity e-", this is the first time I have seen such a symbol "e-".
I have been doing a bit of reading through their site and some of the videos are very interesting (OK I'm a nerd I like this stuff). A CCD with self calibrating Photon amplifiers for low light, it even comes with a Canon EF mount. While you need to cool it to -100oC, this is something that Dr C would enjoy doing.
Unfortunately, their biggest sensor is only 1M pixel.
As for dB, how about a bit of excessive government regulation, so that the users have a say in standards not just industry (or am I dreaming).
Tony
There have been a lot of patents on this type of design, but the only one to make it to market (that I know of) is Foveon, and the chief problem is very poor color accuracy. There is no way to shape the crossover between color channels in the spectral response to more closely match that of the human eye (e.g. red and green wavelengths) -- the same problem that plagues 3-chip video cameras, resulting in significant metameric failure. The other issue is poor SNR in whatever color channel corresponds to the deepest absorption layer (due at least to poor QE there). That said, a lot of people don't mind the grossly inaccurate color of Foveon sensors, so if they could distribute the QE more evenly, it would overcome the terrible low-light reputation Foveon has.Quote:
Originally Posted by ChadS
That dull bump noise followed by the sound of rushing grey matter was my head exploding.