Pages: 1 ... 8 9 [10]   Go Down

Author Topic: interesting article  (Read 89057 times)

bjanes

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 3387
interesting article
« Reply #180 on: June 01, 2006, 09:37:38 am »

Quote
This "exposure" you refer to is a very important one to consider.  It is really the one that is the starting point for determining S/N ratios; once you have chosen this "absolute exposure" as I call it, then the choice of ISO is really counter-intuitive, based on popular belief.  The highest gain-based ISO will give the least noise, unless the camera does a really horrible job of high-ISO amplification.  The only issue with going high with the ISO is clipping (again, assuming absolute exposure is fixed).  The choice of ISO with the manual exposure affects the "digitization depth" or the histogram.
[{POST_SNAPBACK}][/a]

John,

Your post is a bit confusing. By gain I gather you are referring to amplifier gain, which is increased in high ISO settings. CCD system gain is reported in terms of electrons per ADU (alalog to digital unit) and is the inverse of amplifier gain.

While a CCD gain of 1 electron / ADU would be perferable, this would require 16 bit encoding with most current 35mm style cameras, which can record 50,000 electrons easily. The full well of the Canon EOS 1D is ~79,000 electrons and a gain of 1 e/ADU would require 17 bits (inconvient).

While total noise in absolute terms increases with increasing exposure up to full well (according to Poisson sampling), the signal to noise ratio is best at full well. See the discussion by Roger Clark.

[a href=\"http://www.clarkvision.com/imagedetail/evaluation-1d2/index.html]http://www.clarkvision.com/imagedetail/eva...-1d2/index.html[/url]

For optimum S/N one should expose so that the sensor is full well in the highlights and this is related to lux/seconds and not the ISO chosen. If you expose for full well and use an ISO higher than base, clipping will occur with 12 bit AD converters unless the camera has a setting to vary CCD gain, which is usually available only on specialized cameras for scientific purposes. I think that ISO 50 on some Canons sets the gain to 1/2x, and there is a risk of blowing highlights with this setting.

http://www.photomet.com/library_enc_gain.shtml

If you expose at less than full well, then the amplifier gain can be increased. Absolute noise will be reduced, but so will S/N, and this is not what most of us want.
Logged

John Sheehy

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 838
interesting article
« Reply #181 on: June 01, 2006, 02:38:43 pm »

Quote
John,

Your post is a bit confusing.

I don't think so.  My post makes a very reasonable assumption that you are in the real world, and can't come close to saturating the sensor in many situations, without getting motion blur, or shallower DOF than desired.  That is where choosing aperture and shutter speed come in.

Quote
By gain I gather you are referring to amplifier gain, which is increased in high ISO settings. CCD system gain is reported in terms of electrons per ADU (alalog to digital unit) and is the inverse of amplifier gain.

While a CCD gain of 1 electron / ADU would be perferable, this would require 16 bit encoding with most current 35mm style cameras, which can record 50,000 electrons easily. The full well of the Canon EOS 1D is ~79,000 electrons and a gain of 1 e/ADU would require 17 bits (inconvient).

Again, this is the real world.  The current DSLRs can barely deliver 9 or 10 bits worth of really useful information.  A nice, clean 16- or 17-bit readout is a pipedream at this point in time for most people.  Maybe some of the expensive MF backs do better, but base ISO perfomance on most cameras is abysmal.

Also, I think that if in some miracle, a way of reading/digitizing sensors that only includes poisson and dark current noise were made available, that there would be any reason to stop at approximately 1 ADU = 1 electron.  Another bit or two would keep the posterization noise down.  I wish we had that problem.  A clean, 16 or 18-bit digitization of the sensor charges would eliminate the need to change ISOs; ISO 100 under-exposed by 4 stops would be as good or better than ISO 1600 is now, even with the same well capacity.

Quote
While total noise in absolute terms increases with increasing exposure up to full well (according to Poisson sampling), the signal to noise ratio is best at full well. See the discussion by Roger Clark.

I've done this research myself.  I have a spreadsheet with the noise values at various RAW levels at various ISOs from my 20D; I used this data to determine if other people with 20Ds and 30Ds, especially, really had noise problems in their cameras.  In every case it turned out that the noise levels were the same in these cameras, and that it was either exposure or conversion that was emphasizing noise.

Quote
For optimum S/N one should expose so that the sensor is full well in the highlights and this is related to lux/seconds and not the ISO chosen. If you expose for full well and use an ISO higher than base, clipping will occur with 12 bit AD converters unless the camera has a setting to vary CCD gain, which is usually available only on specialized cameras for scientific purposes.

Again, you are assuming that shutter speed and aperture are not real issues.  They are really the most important thing to choose in many situations.  Full exposure at ISO 100 is not an option.  For a lot of wildlife or spontaneous night-shooting in city streets, a full histogram at ISO 1600 isn't even possible.

Quote
I think that ISO 50 on some Canons sets the gain to 1/2x, and there is a risk of blowing highlights with this setting.

ISO 50 is probably implemented as an arithmetic trick in most cases, which loses a stop of highlights, relative to the metering.  In some, ISO 100 doesn't use full-well (Canon 1dmkII), and 50 actually uses a little more of the sensor's DR.

Quote
If you expose at less than full well, then the amplifier gain can be increased. Absolute noise will be reduced, but so will S/N, and this is not what most of us want.
[a href=\"index.php?act=findpost&pid=67092\"][{POST_SNAPBACK}][/a]

No, that's not what we usually want, but that is often unavoidable.

Anyway, you seemed to miss my point - Once you have decided what aperture and shutter speed you want to use (low sensor signal is the primary cause of noise), the *highest* ISO will give the least noise, especially in the shadows (clipping being the only danger).

Let's say you take a shot at ISO 400, and you review the histogram, and notice that there is a stop of unused highlights in the histogram.  You can fill that histogram in two ways; by decreasing the f-stop and/or increasing the exposure time, or by doubling the ISO (some cameras have intermediate f-stops, but they are not always real).  The former will decrease noise, but you might be sacrificing sharpness or DOF, or the lens' sweet spot.  Increasing the ISO will also reduce the noise, not quite as much as increasing the sensor exposure (but more than you might think), but there is no compromise on the Tv and Av settings you want to use.

You seem to be replying to me as if I had written "You get the lowest noise by using the full RAW DR, regardless of ISO", which I did not write.  My original statement was one with a condition - that you have aready chosen the f-stop and shutter speed you want.  There is no option for full-well exposure in many cases of the condition.
Logged

Ray

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 10365
interesting article
« Reply #182 on: June 02, 2006, 12:16:49 am »

Quote
Once you have decided what aperture and shutter speed you want to use (low sensor signal is the primary cause of noise), the *highest* ISO will give the least noise, especially in the shadows (clipping being the only danger).
[a href=\"index.php?act=findpost&pid=67118\"][{POST_SNAPBACK}][/a]


This is a very significant point. I've got lots of ruined shots as a result of giving undue weight to 'absolute' noise in an image. Unfortunately, it's been a slow learning process. I should have used ISO 400 and even 800 more often with my D60. I'm not going to make the same mistake with my 5D. If the shot requires ISO 1600 for an adequate shutter speed and DoF, I'll use it.
« Last Edit: June 02, 2006, 12:18:53 am by Ray »
Logged

John Sheehy

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 838
interesting article
« Reply #183 on: June 02, 2006, 08:49:48 am »

Quote
This is a very significant point. I've got lots of ruined shots as a result of giving undue weight to 'absolute' noise in an image. Unfortunately, it's been a slow learning process. I should have used ISO 400 and even 800 more often with my D60. I'm not going to make the same mistake with my 5D. If the shot requires ISO 1600 for an adequate shutter speed and DoF, I'll use it.
[a href=\"index.php?act=findpost&pid=67150\"][{POST_SNAPBACK}][/a]

With the Canon 20D, 1 stop of under-exposure ("under" means "not to the right", not just "dark") results in more shadow noise than ISO 800 fully exposed.  The 1DmkII is similar, and your 5D probably is, too.

It is rather unfortunate that the majority of digital users believe that "higher ISOs add more noise", and that only a very dark preview image at a low ISO would benefit from a higher ISO with the same absolute exposure.  An image that looks fine in the LCD could easily have 2 stops of unused RAW highlights.

What I long for in future cameras is manual exposure (Av and Tv) with floating ISO, with either a "highlight escape control" that is exponential, like .01%, .1%, 1%, 10%, or a simple EC control as is currently used in auto-exposure modes.  Such a paradigm would be truer to the way a photographer should really make decisions, and the way digital exposure really works.  One of the biggest complaints about such a possible mode is that there is a loss of control of image quality with auto-ISO.  Such complaints are based on a misconception about the role of ISO in noise - ISO only increases noise when it reduces exposure through metering.  Manual Tv and Av with auto-ISO does not have that problem; having the ISO in the viewfinder, optionally flashing at the higher ISOs, would tell the photographer that it is possible to sacrifice f-stop or shutter speed for a cleaner image at a lower ISO.
Logged

bjanes

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 3387
interesting article
« Reply #184 on: June 02, 2006, 10:55:05 am »

Quote
With the Canon 20D, 1 stop of under-exposure ("under" means "not to the right", not just "dark") results in more shadow noise than ISO 800 fully exposed.  The 1DmkII is similar, and your 5D probably is, too.

It is rather unfortunate that the majority of digital users believe that "higher ISOs add more noise", and that only a very dark preview image at a low ISO would benefit from a higher ISO with the same absolute exposure.  An image that looks fine in the LCD could easily have 2 stops of unused RAW highlights.

What I long for in future cameras is manual exposure (Av and Tv) with floating ISO, with either a "highlight escape control" that is exponential, like .01%, .1%, 1%, 10%, or a simple EC control as is currently used in auto-exposure modes.  Such a paradigm would be truer to the way a photographer should really make decisions, and the way digital exposure really works.  One of the biggest complaints about such a possible mode is that there is a loss of control of image quality with auto-ISO.  Such complaints are based on a misconception about the role of ISO in noise - ISO only increases noise when it reduces exposure through metering.  Manual Tv and Av with auto-ISO does not have that problem; having the ISO in the viewfinder, optionally flashing at the higher ISOs, would tell the photographer that it is possible to sacrifice f-stop or shutter speed for a cleaner image at a lower ISO.
[{POST_SNAPBACK}][/a]

The whole noise issue can be summed up by examining the analysis posted by Roger  Clark on his website. His data are for the Canon EOS 1D Mark II.

[a href=\"http://www.clarkvision.com/imagedetail/evaluation-1d2/index.html]http://www.clarkvision.com/imagedetail/eva...-1d2/index.html[/url]

N = (P + r2 + t2)1/2, (eqn 2)

Where N = total noise in electrons, P = number of photons, r = read noise in electrons, and t = thermal noise in electrons
[/b]

Since photon noise is dominant under normal shooting conditions, one should expose so that a maximum number of photons fall on the sensor. If considerations of f/stop and shutter speed preclude exposing to full well, one should give as much exposure as conditions permit. One then has to choose an ISO that will give optimum results.

Current amplifiers introduce very little noise, so from this standpoint, it does not make much difference whether one chooses ISO 100 and boosts the signal in PP or ISO 1600 for exposure to the right, which will, however, make maximum use of the full range of the analog to digital converter and give smaller quantization errors.

Read noise must then be considered. While photon noise is predominant in the highlights, read noise is often predominant in the shadows. One would then choose an ISO that gives the lowest read noise. From Dr. Clark's analysis, ISO 100 has a read noise of 16.6 electrons and ISO 1600 has a read noise of 3.9 electrons. Obviously, the ISO 1600 will not only result in less quantization error but also less read noise, so it is the obvious choice. However, at ISO 3200 read noise increases and dynamic range is reduced, so it is best to stick with ISO 1600. ISO 800's read noise is similar to that of ISO 1600 and ISO 800 is not a bad choice either.

Thermal noise becomes prominent only with very long exposures
Logged

Eric Myrvaagnes

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 22814
  • http://myrvaagnes.com
    • http://myrvaagnes.com
interesting article
« Reply #185 on: June 02, 2006, 11:06:02 am »

Quote
What I long for in future cameras is manual exposure (Av and Tv) with floating ISO, with either a "highlight escape control" that is exponential, like .01%, .1%, 1%, 10%, or a simple EC control as is currently used in auto-exposure modes.  Such a paradigm would be truer to the way a photographer should really make decisions, and the way digital exposure really works.  One of the biggest complaints about such a possible mode is that there is a loss of control of image quality with auto-ISO.  Such complaints are based on a misconception about the role of ISO in noise - ISO only increases noise when it reduces exposure through metering.  Manual Tv and Av with auto-ISO does not have that problem; having the ISO in the viewfinder, optionally flashing at the higher ISOs, would tell the photographer that it is possible to sacrifice f-stop or shutter speed for a cleaner image at a lower ISO.
[a href=\"index.php?act=findpost&pid=67161\"][{POST_SNAPBACK}][/a]
Great idea, John. I second the motion.

But once Canon engineers latch onto the idea, they'll probably want to bury it eleven levels deep in the menu structure, so that it is harder to use than their infamous "mirror lockup" feature.  

Eric
Logged
-Eric Myrvaagnes (visit my website: http://myrvaagnes.com)

John Sheehy

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 838
interesting article
« Reply #186 on: June 02, 2006, 09:56:15 pm »

Quote
N = (P + r2 + t2)1/2, (eqn 2)

Sure that wasn't (P+r2+t2)^(1/2)?  Noise doesn't add linearly.

Quote
Obviously, the ISO 1600 will not only result in less quantization error but also less read noise, so it is the obvious choice. However, at ISO 3200 read noise increases and dynamic range is reduced, so it is best to stick with ISO 1600. ISO 800's read noise is similar to that of ISO 1600 and ISO 800 is not a bad choice either.
[a href=\"index.php?act=findpost&pid=67168\"][{POST_SNAPBACK}][/a]

3200 is non-existent on most cameras, except as an arithmetic trick.  Noise should be double what it is relative to metering, and the same in electrons, as ISO 1600.
Logged

John Sheehy

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 838
interesting article
« Reply #187 on: June 02, 2006, 10:57:00 pm »

Quote
Indeed, and it is not just a matter of noise reduction: underexposing by four stops at ISO 100 means that the signal passes through the last pre-amplifier stages 1/16th as strong, and then has to be amplified 16 times as much in the digital domain to get correct levels. This means that an noise introduced late on the pre-amplifier stage and "quantization noise" from A/D conversion gets amplified by an extra factor of 16.[a href=\"index.php?act=findpost&pid=62976\"][{POST_SNAPBACK}][/a]

I believe that Ray may have been talking about the phenomenon I demonstrated in the pre-pay Galbraith forums.  I took a shot with a wide-angle lens from a dock at dusk of pilings, in the sound, under-exposed for ISO 1600, in manual mode, and then took the same shot dialing the ISO down to 100 without changing the aperture and shutter speed.  I showed the RAW data from both, at the same scale, interpolated, and with WB.  One was 1600 pushed to 10,000, and the other was 100 pushed to 10,000.  The 1600 setting provided the clearest image, by far.  At one time, I would assume that this was due to posterization (1/16 as many levels for the ISO 100), but I decided to posterize the 1600 image so that it had the same number of levels to represent it as the 100 image (20 out of 4096 levels, from 320).  To my surprise, the 1600 image lost almost no image quality whatsoever, when posterized.  It was still orders of magnitude cleaner than the 100 image.  Conclusion - under-exposure at ISO 100 resulted in sloppy images not because of posterization, but because of tremendous readout noise.
Logged

Ray

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 10365
interesting article
« Reply #188 on: June 04, 2006, 10:12:20 pm »

Quote
To my surprise, the 1600 image lost almost no image quality whatsoever, when posterized.  It was still orders of magnitude cleaner than the 100 image.  Conclusion - under-exposure at ISO 100 resulted in sloppy images not because of posterization, but because of tremendous readout noise.
[a href=\"index.php?act=findpost&pid=67232\"][{POST_SNAPBACK}][/a]

John,
We should also bear in mind that noise at base ISO in Canon DSLRs remains fairly constant from model to model, which is reflected in dpreview's noise test charts. If the 1Ds produces less noise (in the final image) at ISO 100 than the older D60, it's only because the 1Ds has a greater number of pixels. On a pixel for pixel basis, D60 noise is actually marginally less.

Canon appears to think that noise at ISO 100 is already good enough (and provided the DR of the scene is not excessive, they are probably right). The early models of Canon DSLRs, including the D30 and 1D, did not show this 'lower read-out noise' at higher ISOs. The 10D set the trend to reduced noise at higher ISOs and it's just got better with successive models.

Are you sure that reduced read-out noise is the only factor here? If so, what are the obstacles to reducing read noise at ISO 100?
Logged
Pages: 1 ... 8 9 [10]   Go Up