What does changing the ISO do, exactly?

[DannoPiano]

Good find, BJL. I think I have a fairly good grasp for ISO in the digital realm. But I wonder if we have reached the limit of what we can know without delving into trade secrets of "exactly how". I'm satisified with the knowledge I have gained so far because it allows me to understand how it affects taking pictures and using other ideas like noise vs. grain and dynamic range. Anything beyond this probably won't help me take better pictures! While it is nice to know how stuff works, I can't remember anyone ever having a discussion about how to make film at different ISO's. That was left to film makers - now it seems everyone wants to know how it all works and get their 2 cents in. Times are changing, I guess.

[Steve Kerman]

Sorry about the duplicate post.  If someone has the ability to delete this copy of the thread, please do so.  Thanks.

[DannoPiano]

At first I was inclined to agree with you, samirkharusi. But you have to remember that ISO is a measurement of Illuminance and time. CMOS sensors are capable of detecting as little as one photon of light, the same as your eye. Over an average exposure of 1/30 sec., how many photons will hit any given photon sensor on the chip? The answer is probably alot more than you need to get some kind of reading. (Is there a mathematical formula to determine this?) But, if you want an accurate reading you have to take many samples. It's a matter of statisictics. The ADC converts to 8/12/16-bit luminance values, but I would guess that there are probably alot more photons hitting a given sensor than there are bit values in a typical exposure. The more values the ADC has to work with, the more accurate its representation can be, similar to taking a statistical sample in a public opinion poll. The more samples you take, the more accurate your estimate will be. And my biggest point here is that the photon capture process is an estimate of what really happened, storing an exact count of every photon that struck a sensor would consume too much memory.

But don't confuse the signal amplifier with the ADC process. The amp just gets those signals up to a voltage level the converter can use to make the transistors work (probably 3.3v).

With that in mind, I don't see why you couldn't simulate ISO settings by taking smaller samples (shorter time) of photons. The problem then becomes that with smaller samples the more likely you are to have "noise" or errors in your samples. That's where anti-noise software comes in, to smooth out the bumps and quirks in your sample. In a sense, it cheats by making "noisy" pixels look like their neighbors which gives the appearance of a clean photo when the data was probably not there to begin with.

Does this decrease dynamic range? I would say no if the average sensor sample is greater than the maximum bit value and the ADC is converting down to the smaller values. If the average sample was less than the maximum bit value, then yes you would have decreased dynamic range because you're not capturing all of the potential information. In addition, noise could be "louder" than the original signal distorting the image completely.

So I guess the question to ask is: Are the luminance values produced by the ADC
a) lower than the average # of photons striking a sensor area
 one-to-one correspondence of # of photons striking a sensor area
c) greater than the average # of photons striking a sensor area

Is it possible that in the same exposore that two or even all three of these conditions occur?

P.S. I could be wrong about all of this, what do I know? The ADC converts to 8 bit values? Or is 12? or 16? The principle I think is the same. I've been doing my darndest to find info on this online and this is my educated guess from what I have learned so far. I'm entirely open to correction, since I'm trying to determine what actually happens. Will camera makers ever tell us, or is this a dark secret?

[jwarthman]

Ray,

Correct exposure for optimum, noise-free results is at one ISO only. Increase the ISO and you are basically sending the camera a forewarning, "I'm about to send you an underexposed image. Make the best of it." At which point, all the electronic wizadry the camera can muster whirs into action to often produce a surprisingly good result.

You make some good points! Even so, I still prefer to think of the ISO setting on a digital camera as analogous to the size of the bucket. If I have a small bucket, I can't add much water (light) before it overflows! Conversely, if I have a large bucket, the same amount of water (light) will leave some of the bucket capacity unused (e.g. an underexposed image).

Notice that I'm not claiming that ISO is independent of noise! I agree that one ISO setting probably represents some "native" sensitivity of the sensor/amplifier combination. (Although I don't think "noise-free" is a fair characterization.    ) It seems that we both agree that increasing the ISO from this "optimum" is done by some analog and/or digital magic within the camera, with the introduction of noise. Others have suggested this might also reduce dynamic range.

get back to the bucket analogy, I would say the size of the bucket equates to the size of the photodetectors, the # of photons to the quantity of water, the diameter of the hose to the diameter of the aperture, the ISO to the inverse of the water pressure and the shutter speed to the time the water is flowing.

I prefer to cleanly separate the environment that brings light (water) to the sensor (bucket) from the sensitivity of the sensor. The water flow into the bucket is analagous to the light striking the sensor/film. The water pressure is, I suppose, part of that system - it determines how much water enters the bucket, and how fast. As far as the analogy goes, I guess water pressure would be equivalent to the brightness of the scene. The sensitivity of the sensor/film is something apart from the water pressure. To use your concept, perhaps the bucket size is a function of both native sensor sensitivity, and the amplification used to increase the ISO beyond the native sensitivity - which also increases the noise.

Does this make sense?

Enjoy!

-- Jim

[jwarthman]

This is a great discussion - thanks for the interaction!

Ray,
Was it red or white wine?  ::

I do believe you're right - if you equate "bucket" with just the photosite. But remember, my analogy was intended to describe the sensor "system", not just the photosite:

The size of the bucket is analogous to the sensitivity of the sensor (including the photodiode and amplifier) - the ISO, if you will.

(Actually, I should also have included the A/D conversion step in my description.) When I say "bucket size", I don't mean to imply the capacity of the photosite.

Bruce,
I think you've brought clarity to the ISO issue! Your analogies address both the fixed capacity of the photosite, as well as the variable ISO by the time A/D conversion is done. I like it!    


Danno,
Is this making sense - without introducing a "time" parameter?


Enjoy!

-- Jim

[BJL]

Danno,

   thanks for the Roper Scientific link. I do not think it is quite true that increased ISO is achieved without loss of dynamic range though; reading carefully, they describe options of "high sensitivity" [given by the on-chip gain method] OR "wide dynamic range", not both at once.

To understand the options, you have to take account of four main stages where noise (or error in numerical values) are introduced.

a) "Shot noise": random fluctuations in the number of photons detected by the sensor, due to fundamental physics; there is nothing to be done about that except using bigger photo-sites, so this will some day be the ultimate limit on low light sensitivity at a given photo-site size.

 "Dark current": thermally caused random changes in electron counts; this seems to be one stage where progress is making small photo-sites work better, perhaps by operating at ever lower voltages and hence lower temperatures.

c) "Read noise": noise introduced in the output amplifier between the sensor chip and the A/D convertor (probably another place that progress is happening.)

d) "Discretization error": rounding off to the nearest integer value in the digital output. But with 12 bit or better A/D conversion, this is small relative to earlier sources in current sensors, so I will mostly ignore it from now on.


The first two categories of "in-sensor noise" fundamentally limit the dynamic range, which cannot exceed the ratio between this noise level and the maximum countable number of electrons: any gain or amplification also increases this part of the noise. If you push the ISO, the maximum countable number of electrons is decreased (in my analogy, it take less electrons to fill the more slender graduated cylinder), so some dynamic range is lost.

At very low light levels, dark current dominates over shot noise (shot noise is lower at photosites receiving less light, while dark current is not, and instead is proportional to exposure time: sensors with good long exposure noise probably have low dark current). So dark current is the main villain in deep shadow noise produced in the sensor itself.

On the other hand, the technique of on-chip gain described at Roper Scientific happens before the output amplifier, so avoids the amplification of read noise (and discretization noise.)

So the degree of loss of dynamic range during "pushing" depends mostly on the relative size of dark current noise vs read noise.


P. S. I am fairly sure that something similar to on-chip gain is an important part of Canon's being able to produce CMOS sensors that reverse the high noise reputation of earlier CMOS sensors.

[Ray]

There are quite a few sites that offer a basic primer on CCD and CMOS processes, but I can't find anything relating specifically to the method by which changing the ISO setting on the camera increases the gain or amplification of the charge that has accumulated in each photosite as a result of photons knocking off electrons.

I would still maintain that the 'sensitivity' of the 'electron collecting' site is not changed by changing the ISO setting on the camera, but rather there is extra amplification applied to each photodetector's charge at the read-out stage. Some of this might even be done before A/D conversion.

As BJL has pointed out, it's difficult to amplify a signal without also amplifying at least some of the noise, just as for example, when you receive a weak signal on your radio, turn up the volume and get a lot of background crackle. This is called poor signal-to-noise. The options are, either increase the strength of the signal (install a better antenna) or devise some clever trick of reducing the noise.

Increasing the ISO setting on the camera effectively reduces the signal strength through the process of letting less light in. There are consequently fewer photons striking each photosite, fewer electrons that are knocked off the silicon crystals and a smaller charge than would otherwise be the case if the ISO setting were optimal (100 ISO).

The trouble with the bucket analogy is that this term is often used to describe the 'electron collecting' capacity of each photosite. Terms like 'full well' are used to describe a photodetector that can carry no further charge. Getting the wells full with respect to the highlight areas of the scene being photographed, is all part of the recommended process of exposing to the right of the histogram. This will ensure maximum dynamic range and maximum S/N and this can only be achieved if at least some of the buckets are full at the collection stage (not at some later transferral and processing stage).

Increasing the ISO setting is therefore bound to reduce both S/N and D/R. I try to avoid it

[BJL]

Ray,

    thanks for digging out those guidelines of S/N>40 for exellent image quality, >10 for aceptable, they at least set upper limits on dynamic range at various ISO's.

For example, to get above S?N of 40, you have to count 40 squared = 1600 electrons just to get above shot noise. Other noise sources are comfortably less than 40 electrons under normal circumstances I believe, so you are probably OK at that level. Digging one more number out of Kodak's site, of about 50,000 electrons as the maximum that several of their DSLR sensors can count, and you have a typical dynamic range of 1600 to 50,000, a ratio of 30 or five stops; (close to transparency film?).

Probably in shadows that are more than five stops from the top (which are even hard to print without compressing contrast), the more liberal "acceptable level" is often all you can make out, and that lets you go down to 100 electrons, which suggests a dynamic range of 9 stops. This at least fits with figures in the range 8 to 9 stops I have seen is some technical articles that, however, are vague about their definitions of dynamic range.

Every stop higher ISO loses one stop at the top of the range, so loses one stop of dynamic range for high contrast subjects.

As to when ISO pushing is applied, it seems from the Roper Scientific reference that the traditional place is gain in the amplifier feeding the A/D converter, with some sensors also doing something with "on-chip" gain; in either case, before A/D conversion. I have not heard any mention of any camera doing it in the digital domain (despite my earlier speculations).

[Ray]

I see relevance for owners of the Kodak 14n in these D60 results. Since Breezbrowser does such a good job in exposure compensation (and Capture One possibly even better) it would be of little concern to me if the D60 had only the one ISO setting of 100.

Now I don't recall the D60 being criticised for being unacceptably noisy at ISO settings above the base of 100. And as I remember, at ISO 80 the 14n is as noise-free as the 1Ds. Therefore I see no reason why software processing should not be able to produce acceptable results at ISO 400 and even 800. Not sure about the long exposure limitation, though.