Page 1 of 1

Do "Standard" ISOs Have Less Noise?

Posted: Sun Sep 28, 2008 8:59 am
by ebkw
Some time ago in this thread

http://www.naturescapes.net/phpBB3/view ... native+ISO

There was a discussion of the 30D and which ISO has less noise. I thought there was another another one that said that standard ISOs, 100, 200, 400, 800, and 1600, had less noise than the inbetween ones. I am trying to find that thread and find out if that is really true or if I imagined the whole thing?

I have been keeping away from using the inbetween 1/2 ISOs because of this. Am I wrong?

I am using both MK2N and 40D bodies now.

Posted: Sun Sep 28, 2008 10:08 am
by ejmartin
It is a somewhat complicated issue.

The 40D, like the 30D before it, implements intermediate ISO's by software multiplication after digitization of the raw data. So there aren't really intermediate ISO's, as far as the camera electronics is concerned; instead, the camera overexposes by 1/3 stop at ISO 160-320-640 etc, and underexposes by 1/3 stop at ISO 125-250-500-1000 etc., using the nearest "main" ISO 100-200-400 etc. That is, it internally only has the main ISO's; it adjusts the metering by 1/3 stop; and then after the capture it multiplies or divides the raw values by 1.25 to make the data look as though it came from the indicated ISO.

What does this do for noise? Well, exposing to the right (ETTR) is a technique for reducing noise, because to push the histogram to the right you must increase the light gathered, and the more light gathered the less apparent the noise is relative to the signal. Similarly, underexposing increases noise by making the signal stand out less above the noise. What the camera is doing at ISO 160-320-640 etc is overexposing by 1/3 stop at ISO 200-400-800 etc and dividing by 1.25 after the fact. But this also removes 1/3 stop of highlight headroom, because internally the camera uses a higher ISO (200, say) that amplifies the signal more than (say) ISO 160 and so pushes 1/3 stop of highlights past the saturation point of the electronics (the A/D converter, specifically) which are then lost. Dividing by 1.25 will not bring them back since they are already clipped.

Conversely, the other set of intermediate ISO's (125-250-500 etc) underexpose the image by 1/3 stop and then push the exposure by 1/3 stop before writing the raw data. Like any underexposure, this increases the noise.

Bottom line: on the 40D, ISO 125-250-500 etc are underexposing at ISO 100-200-400 etc and are no different than using ISO 100-200-400 etc and manually dialing in -1/3 EV of exposure compensation. ISO 160-320-640 etc are overexposing at ISO 200-400-800 etc, and are no different than using ISO 200-400-800 etc and manually dialing in +1/3 EV of exposure compensation. The only difference will be that you will have to apply appropriate exposure adjustment during raw conversion if you stick to the main ISO's, if you use the intermediate ones the camera will do it for you internally. I personally use the custom function on my 40D that turns off the intermediate ISO's; I prefer to set exposure compensation myself, knowing that I have the same amount of highlight headroom available at each main ISO, rather than having to remember that I have an extra 1/3 stop at some ISO's and 1/3 stop less at some others, etc, when using the intermediate ISO's.

Now for the 1D2n. In its pro series cameras (including the 5D and presumably the 5D2), Canon does use hardware amplification to implement the intermediate ISO's, rather than using software multiplication of the raw data after the fact as in the 30D/40D (and presumably the 50D as well). However, the implementation is by a separate amplifier that amplifies the signal before digitization, using the output of a previous stage of amplification at the next lowest main ISO. For instance, both ISO 250 and 320 are derived from an additional stage of amplification from an earlier stage of amplification at ISO 200. However, this extra amplifier is so noisy, that it provides little or no advantage over underexposing by 1/3 or 2/3 stop at the next lower main ISO -- there is little or no gain noise-wise (essentially none above ISO 400, and very little below ISO 400), but there is again a loss of 1/3 or 2/3 stop of highlight headroom due to the extra amplification pushing highlights past the saturation point of the A/D converter. And again, I have the custom function set on my 1D3 that turns off the intermediate ISO's; I would prefer to have the extra highlight headroom of underexposing at the next lower ISO, if I need the faster shutter that the higher intermediate ISO meters for.

And for all you Nikon users, the D300, D3 and D700 all implement intermediate ISO's as hardware amplification in a single stage; there are no drawbacks to using them, no convoluted chain of logic needed 8)

Posted: Sun Sep 28, 2008 10:20 am
by Eric Chan
Emil, I have not noticed a loss of 1/3 or 2/3 stop of highlight headroom at the intermediate ISOs on the 1-series cameras.

How did you determine this -- by examining saturated white levels?

Posted: Sun Sep 28, 2008 10:30 am
by ebkw
Thanks, Emil! I will turn off the 1/3 stops on both bodies. That simplifies counting stops when I can't see the window.

Re:

Posted: Sun Sep 28, 2008 11:09 am
by ejmartin
madmanchan wrote:Emil, I have not noticed a loss of 1/3 or 2/3 stop of highlight headroom at the intermediate ISOs on the 1-series cameras.

How did you determine this -- by examining saturated white levels?
Sorry, I mis-spoke (mis-wrote?) slightly; but the thrust of what I stated is correct, it just needs a bit more explanation for people to implement the conclusions properly in the field.

There is no loss of highlight headroom on the 1 series with respect to metering of 18% (or 12.7%) gray; middle gray is the same number of stops below raw saturation for all ISO. What is true is that, for a given series of ISO's 200-250-320, or 400-500-640, or 800-1000-1250, the read noise is essentially constant in photoelectrons. The saturation level of the raw data is however decreasing by a factor 1.25 in terms of photoelectrons, for each 1/3 stop increase in ISO. So if the noise floor is constant in terms of photons but the saturation level is decreasing, dynamic range is reduced (by 1/3 stop from ISO 200 to 250, and by another 1/3 stop from 250 to 320; similarly for the other series). Now, let's see what to do with that extra DR afforded by the next lowest main ISO.

Consider two shooting scenarios:

1. Set the camera to ISO 320 and expose according to the indicated metering on the camera.
2. Set the camera to ISO 200 and underexpose by 2/3 stop according to the metering on the camera; in other words, the camera exposes at the same shutter speed and aperture -- the same absolute exposure -- as scenario 1.

Choice 2 will have the same amount of shadow noise, because the read noise is essentially identical in terms of photon equivalents and so the shadow S/N ratio is the same. But because scenario 2 is underexposed 2/3 stop relative to metering, there is an additional 2/3 stop of headroom in highlights. The image will of course be 2/3 stop darker when brought into the raw converter; just bring it up 2/3 stop with the exposure slider (hence performing an amplification in software post-capture), the result will be no different than that extra stage of hardware amplification in the camera did (yes the level spacing will be 1.6 times wider as a result, but the noise is well above this level and so it makes no difference).

Re:

Posted: Sun Sep 28, 2008 11:51 am
by Jens Peermann
ebkw wrote:Thanks, Emil! I will turn off the 1/3 stops on both bodies. That simplifies counting stops when I can't see the window.
I don't think there is such a function on the 40D. At least I cant find it in the manual.

Posted: Sun Sep 28, 2008 12:07 pm
by Scott Linstead
Emil, once again I am astounded with the wealth of knowledge that you possess on the theory of digital imaging. Thanks for the enlightenment, I had no idea that there was different ways that a manufacturer could approach intermediate ISOs.

Re: Re:

Posted: Sun Sep 28, 2008 2:44 pm
by ejmartin
Jens Peermann wrote:
ebkw wrote:Thanks, Emil! I will turn off the 1/3 stops on both bodies. That simplifies counting stops when I can't see the window.
I don't think there is such a function on the 40D. At least I cant find it in the manual.
It's custom function I-2. Switch it from 0 (1/3 stop ISO increments, the default) to 1 (1-stop increments).

Just to sum up, what I'm saying is

1. For the 40D, rather than using the intermediate ISO's, one does as well or better going to the nearest main ISO and dialing in exposure compensation of +/- 1/3 EV. There is no difference, but you are consciously aware of how much highlight headroom you have available.

2. For the 1 series, rather than use the intermediate ISO's, one does better by going to the next lower main ISO, and dialing in EC of -1/3 or -2/3 EV; this will give the same shutter speed/aperture and noise performance, but provide 1/3 or 2/3 stop more dynamic range (more headroom in highlights relative to the use of the intermediate ISO's). ISO 125 and 160 have a marginal benefit in shadow noise relative to underexposure at ISO 100, but it's quite small and to my mind not worth the loss of overall DR.

And again, this is only for Canons; for Nikon the implementation of intermediate ISO's is done better and one should feel free to dial them in as needed.

Posted: Sun Sep 28, 2008 2:44 pm
by Eric Chan
emil, I guess what I'm asking you is that if you examine a 1D2n's saturation level (on a 12-bit scale) does an ISO 320 image clip at a different point than an ISO 400 image?

Re:

Posted: Sun Sep 28, 2008 2:49 pm
by ejmartin
madmanchan wrote:emil, I guess what I'm asking you is that if you examine a 1D2n's saturation level (on a 12-bit scale) does an ISO 320 image clip at a different point than an ISO 400 image?
Relative to what? I can't answer the question until I know what you want to hold fixed:

The saturation point in raw levels?
The saturation point in photoelectrons?
The saturation point in stops above metered middle grey?
The saturation point in stops above the noise floor?

They are all different, except perhaps the first and third.

Posted: Sun Sep 28, 2008 2:50 pm
by Eric Chan
The 12-bit raw level from a mosaiced image.

Re:

Posted: Sun Sep 28, 2008 2:57 pm
by ejmartin
madmanchan wrote:The 12-bit raw level from a mosaiced image.
That is more or less the same for ISO 320 or 400. However, the number of photons/raw level is 1.25 times higher at ISO 320, so that saturation raw level (relative to the black point of 128 on a 1D2n) represents 25% more photons. The raw level at ISO 320 that represents the same number of photons as the clipping level of ISO 400, is .8 of the clipping level. That is why, if the read noise level is the same in photoelectrons, one will have more headroom by underexposing at the lower ISO while having the same performance in shadows.

Posted: Sun Sep 28, 2008 3:02 pm
by Eric Chan
Thanks emil. And I assume you determined the photos/raw level difference for ISO 320 vs 400 (as well as the whole series) by measuring the gain factor separately for each ISO?

Re:

Posted: Sun Sep 28, 2008 3:27 pm
by ejmartin
madmanchan wrote:Thanks emil. And I assume you determined the photons/raw level difference for ISO 320 vs 400 (as well as the whole series) by measuring the gain factor separately for each ISO?
I haven't actually done that particular measurement. I have seen data for the main ISO's, and I do know that the intermediate ISO's on the 1 series are implemented in hardware. I suppose it is logically possible that the intermediate ISO gains are not calibrated properly relative to the main ISO's, and that 3rd party raw converters are in on the conspiracy and apply different EC for the intermediate ISO's during raw conversion, so that nobody has noticed the difference yet. :wink:

Re:

Posted: Sun Sep 28, 2008 3:35 pm
by ejmartin
Scott Linstead wrote:Emil, once again I am astounded with the wealth of knowledge that you possess on the theory of digital imaging. Thanks for the enlightenment, I had no idea that there was different ways that a manufacturer could approach intermediate ISOs.
Pshaw :oops: :)

Just to add one more option to the ones discussed above, some medium format backs implement ISO entirely in software; ISO is written into metadata, specifying the amount of exposure compensation the raw converter should apply while converting the image, and the camera itself has only a single hardware amplification applied to the sensor data.

Re: Re:

Posted: Sun Sep 28, 2008 4:14 pm
by Jens Peermann
ejmartin wrote:
It's custom function I-2. Switch it from 0 (1/3 stop ISO increments, the default) to 1 (1-stop increments).
Found it! It's always been there, I just didn't look close enough and thought it offers increments in 1/3 and 1/2 stop, just like the exposure compensation. Thanks for opening my eyes.

Posted: Sun Sep 28, 2008 8:54 pm
by E.J. Peiker
By far the best explanation of the Canon noise charachteristics on in between ISO's I've seen. This should be turned into an article! :)

Posted: Mon Sep 29, 2008 6:49 am
by Kari Post
Wow thats a mouthful. Emil, thanks for the detailed explanation. I think I'll just keep my camera as is :)

Posted: Mon Sep 29, 2008 8:39 am
by ColorChange
Great job.