RE: Understanding Sensor Issues

[JRT]

The August column makes some good points.  However, in stating flatly that bigger sensors are better, other factors are not considered.

The main additional consideration is the aperture of the lens being used.

Smaller formats can have larger lenses and you can increase the signal/noise ratio by increasing the signal (more light on the sensor) as well as decreasing the noise (larger sensor elements).

A 1 inch video format sensor (size = 16mm movie, 12.8 x 9.6 mm) can have 12 MegaPixels with a 3.125 micron sensor element pitch.  For that format, a 16 mm f/1.4 lens would not be really fast and a f/0.95 is certainly possible (25mm f/0.95 C-mount movie lenses are generally available -- they are not fancy designs, just a standard (Planar) type design with element splitting).

Yes, an APS size sensor with the same resolution would have sensor elements with about four times the area (3.845 x).  This could mean about twice the signal to noise ratio with the same amount of illumination (but would probably be a bit less).  But, it is possible with the smaller format to provide 2 stops more illumination (easy to provide 1 stop more).  A 30mm f/0.95 lens is possible at a price, but how would the cost of a 16mm f/1.4 lens (1" video format) compare with a 30 mm f/2.8 APS lens?

It appears to me that this is valid in larger formats as well where a normal lens would probably be only f/2.8 or less.

The other issue is diffraction.  IIRC, the actual size of the Airy Disc is determined only by the F number.  So, bigger is better with the same aperture..  But an aperture 2 stops larger and it is half the size.  So, the photographic effect is about the same (16mm vs. APS).

[Ray]

It has always been the case that lenses out-resolve film or sensor. The difficulty that sensors have in resolving detail from lenses is partly due to the inability of pixels to record really low contrat (low MTF) detail because of noise of various types and lack of sufficient 'quantum efficiency' of the photodiodes.,

Sensors have a theoretical maximim resolution depending on the pixel pitch of the sensor, but an over all maximum quality depending on pixel size, pixel pitch, sensor design, noise reduction technology etc., etc.

The effects of diffraction also relate to sensor size (not pixel size as is so often stated by confused people).

(I'm being provocative here, because pixel size does indirectly relate to sensor size   ).

[John Sheehy]

Yes, an APS size sensor with the same resolution would have sensor elements with about four times the area (3.845 x).  This could mean about four times the signal to noise ratio with the same amount of illumination (but would probably be a bit less).
[a href=\"index.php?act=findpost&pid=72930\"][{POST_SNAPBACK}][/a]

That would be a maximum S/N ratio with 2x improvement, not 4.  The maximum S/N is proportional to the square root of the number of photons collected, as far as shot noise is concerned.  The kind of noise that limits DR at low ISOs is mainly readout noise, which has no direct link to pixel size.  Bigger wells translates to lower native sensor ISO, and less shot noise at all ISOs, since more photons are collected per sample.  Shot noise shows mainly in the highlights of low ISOs, and the highlights and midtones of high ISOs.  Bigger pixels only help noise and DR at the camera's lowest ISO infinitessimally, because readout noise dominates.

[JRT]

The effects of diffraction also relate to sensor size (not pixel size as is so often stated by confused people).

(I'm being provocative here, because pixel size does indirectly relate to sensor size   ).
[a href=\"index.php?act=findpost&pid=72933\"][{POST_SNAPBACK}][/a]

Actually not.  Simple approximation is that diffraction starts to degrade the image when the Airy disc is equal to the pixel pitch.

There is a much more complicated way to calculate this where depth of detail is calculated using the geometric sum of the effective circle of confusion and the diameter of the Airy disc.  With a film camera, the resolution of the film can also be considered in the calculation.  This method helps to explain the apparent increase in the depth of field with low resolution digital sensors.

[JRT]

That would be a maximum S/N ratio with 2x improvement, not 4.  The maximum S/N is proportional to the square root of the number of photons collected, as far as shot noise is concerned.  The kind of noise that limits DR at low ISOs is mainly readout noise, which has no direct link to pixel size.  Bigger wells translates to lower native sensor ISO, and less shot noise at all ISOs, since more photons are collected per sample.  Shot noise shows mainly in the highlights of low ISOs, and the highlights and midtones of high ISOs.  Bigger pixels only help noise and DR at the camera's lowest ISO infinitessimally, because readout noise dominates.
[a href=\"index.php?act=findpost&pid=72951\"][{POST_SNAPBACK}][/a]

Yes, if shot noise is dominant then that is true.  Actually I know that and appear to have gotten my posting confused.  Will correct it.

The readout noise issue is a problem with my assertion since you can't simply add lower ISOs of 25 & 50 to a smaller sensor with a larger lens.  

[bjanes]

That would be a maximum S/N ratio with 2x improvement, not 4.  The maximum S/N is proportional to the square root of the number of photons collected, as far as shot noise is concerned.  The kind of noise that limits DR at low ISOs is mainly readout noise, which has no direct link to pixel size.  Bigger wells translates to lower native sensor ISO, and less shot noise at all ISOs, since more photons are collected per sample.  Shot noise shows mainly in the highlights of low ISOs, and the highlights and midtones of high ISOs.  Bigger pixels only help noise and DR at the camera's lowest ISO infinitessimally, because readout noise dominates.
[{POST_SNAPBACK}][/a]

John,

Your statement about bigger pixels not affecting read noise is incorrect. It is true that the read noise expressed in electrons is similar for large and small pixels. However, when a given read noise in electrons is converted to a pixel data number using the gain factor, the effect of pixel size is large. See Table 3 of Roger Clark's  analysis:

[a href=\"http://www.clarkvision.com/imagedetail/does.pixel.size.matter/index.html]http://www.clarkvision.com/imagedetail/doe...tter/index.html[/url]

"The realization from Table 3 is even though the read noise is similar in terms of electrons, the effect on the image is huge because of the gain factor. Thus shadow detail on a small sensor is severely compromised as the gain factor drops."

[John Sheehy]

Your statement about bigger pixels not affecting read noise is incorrect.

I don't think so; not in the intended context - in the shadows, at the noise floor, affecting DR.  Shot noise is not read noise.  It is actually signal, just not the kind of signal we want, so we call it noise.  It isn't generated by the system, it is the very fabric of the system.

It is true that the read noise expressed in electrons is similar for large and small pixels. However, when a given read noise in electrons is converted to a pixel data number using the gain factor, the effect of pixel size is large. See Table 3 of Roger Clark's  analysis:

http://www.clarkvision.com/imagedetail/doe...tter/index.html

"The realization from Table 3 is even though the read noise is similar in terms of electrons, the effect on the image is huge because of the gain factor. Thus shadow detail on a small sensor is severely compromised as the gain factor drops."
[a href=\"index.php?act=findpost&pid=77349\"][{POST_SNAPBACK}][/a]

Read noise is not really "electrons" in the context of counted electrons.  Read noise is converted to electrons as a unit, to facilitate a correlation between photon count (and its inherent shot noise) and readout noise.  Both could also be expressed in terms of RAW values.  The former is more useful for comparing absolute exposure at different ISOs (where ISO has no affect on photon count and noise, only on readout noise), and the latter is better for talking about relative (histogram) exposure at any given ISO.

The fact is, all Canon cameras are limited in RAW exposure to 3967 levels (20D) or less (5D with 3524) above blackpoint, and in that context, the readout noise is very similar at any given ISO on various models; generally within 5% (except for fabricated ISOs, not based on purely optimized amplification).  This means that the "noise floor" (I put that in quotes because I don't like some of the connotations it gives), at the pixel level, is very close, relative to maximum signal, at the same ISO, on the 20D, 30D, 5D, and 1DmkII (excluding the dodgey in-between ISOs).  I haven't examined 1DsmkII data, but I would imagine that the results are similar.

You can also see from the Table 3 that you referenced, that the 1DMkII doesn't even use full well at any ISO; it uses 52,300 for a saturated ISO 100, but has a full well capacity of 79,900 (making the native ISO about 65.5).  This means that dynamic range would be reduced by about 34.5% if the camera implemented ISO 50 with half the amplification of ISO 100.  Also, if the camera actually had an optimized ISO 64 amplification, it would most likely have the same limit to DR as ISO 100; a flatline level of readout noise unrelated to amplification, seen in the almost indistinguishable readout noise at ISOs 100 and 200 in recent Canon cameras.

The number of photons collected at ISO 100 on the 1DmkII (52,300) is actually less than the Canon 10D (55,000).  The big pixels on the 5D and 1DmkII really have very little RAW-noise-related benefit; their real benefit is that they allow you to use more of the focal plane, to get more resolution out of a lens, and use more pixels for the capture.  The result is that less sharpening is needed for high-frequency content (high frequency within the image, not within a mm of focal plane; an APS-sized crop from the FF 5D or 1.3x 1DmkII with the same lens and focal length would have less resolution, but more pixel-level sharpness).

[bjanes]

I don't think so; not in the intended context - in the shadows, at the noise floor, affecting DR.  Shot noise is not read noise. 
[a href=\"index.php?act=findpost&pid=77398\"][{POST_SNAPBACK}][/a]

Despite all your obfuscations, large pixels have better signal to noise both in the highlights and shadows. Look at Roger's graph 3b. As it turns out, the 1D MII is photon limited at both highlight and shadows of a normal photographic image, and read noise is not a significant factor.

The Canon S70 point and shoot consumer camera is photon noise limited for all levels above about 10 photons.

These equations excerpted from Roger's post summarize noise characteristics. In the case of a small sensor with a small gain, the effect of read noise on the pixel data numbers is magnified when one divides by the gain.

The noise from the sensor in electrons is, to a good approximation:

N = (#electrons^2 + read_noise^2)^1/2

But we are concerned with noise in our images, so the 16-bit integer scaled noise is:

N16b = {([(DN*fw/65535)^1/2] * 65535/fw)^2 + (read_noise * 16/gain)^2}^1/2

where N16b is the 16-bit noise in image data numbers (DN), fw is the pixel full well capacity in electrons, read_noise is the read noise (e.g. from Table 2) in electrons, and the gain is the gain in electrons/DN (e.g. from Table 2).

[John Sheehy]

Despite all your obfuscations,

Obfuscations?  Hardly.  My style is to get right to the heart of what really matters.  In this context, my interest is what it is that limits DR, and if there is any benefit to DR by having slightly larger pixels, as the 5D and 1DmkII do, compared to the APS-sized Canons like the 20D.  The general point is that larger pixels do not do much to increase RAW DR, if read noise is present (a bad converter, of course, could black out at a level where photon noise is still most significant).  The Canon-specific point is that Canon doesn't even use the full-well capacity in its larger-sensor DSLRs efficiently.  The pixels are geometrically bigger, with some optical advantages, but not much pixel-level noise advantage.

large pixels have better signal to noise both in the highlights and shadows. Look at Roger's graph 3b. As it turns out, the 1D MII is photon limited at both highlight and shadows of a normal photographic image, and read noise is not a significant factor.

Not surprising, since he only charted areas affected mainly by photon shot noise on the 1DmkII.  Had he gone a few steps lower in exposure, and had the S/N axis gone lower, too, you would see all the lines curving down, the lower ISOs faster.  This chart is not for determining DR, of course; DR is the ratio of maximum signal to the lowest usable signal, or the noise floor (readout noise), in the same units.
He is comparing tiny-pixel cameras to large-pixel cameras in this chart.  I thought we were talking about 5D and 1DmkII-sized pixels to 8MP APS-sized pixels.  That was what all my comments were about, anyway.  Had he charted the 1DmkII against the 20D, the lines from each camera, at the same ISO, would be hugging each other, and *THAT* is my point.

I am kind of skeptical about his figure of 16.6-electron readout noise at ISO 100, though. (16.6/52300)*3590 (typical1DmkII RAW clipping point, adjusted for blackpoint) = 1.14.    That would mean that the standard deviation of a RAW blackframe would be about 1.14, which is much lower than I have ever heard for a Canon.  Anyone with a 1DmkII and a PC with the program IRIS can verify the blackframe noise.  Just shoot at ISO 100, a very short exposure, lens cap on, and load the RAW into IRIS with the "Load a RAW file" menu item.  Then, open the command console, and type "stat".  My guess is that the Sigma value will be somewhere in the 1.8 to 2.0 vicinity.  If it is really 1.13 or so, then this camera should be the clear DR leader among the Recent Canons.

Roger's ISO 100 capacity for the 10D is not correct, as far as I remember; it is 55,000 for the 10D.  The figure presented is correct for the 20D, although it doesn't halve at ISO 200, as 100 is missing about 0.25 stops of highlight headroom.

The Canon S70 point and shoot consumer camera is photon noise limited for all levels above about 10 photons.

It is not surprising that read noise is significant only at lower exposure levels on the S70; it has 4300 electrons at RAW saturation at ISO 100; that is as many maximum photons as would be collected on a 1DmkII if it had an ISO 1216 (100*52300/4300).  Comparing the 1DmkII with the 20D, however, the 1DmkII would collect as many max photons at ISO 102.5 as the 20D does at ISO 100.

These equations excerpted from Roger's post summarize noise characteristics. In the case of a small sensor with a small gain, the effect of read noise on the pixel data numbers is magnified when one divides by the gain.

The noise from the sensor in electrons is, to a good approximation:

N = (#electrons^2 + read_noise^2)^1/2

But we are concerned with noise in our images, so the 16-bit integer scaled noise is:

N16b = {([(DN*fw/65535)^1/2] * 65535/fw)^2 + (read_noise * 16/gain)^2}^1/2

where N16b is the 16-bit noise in image data numbers (DN), fw is the pixel full well capacity in electrons, read_noise is the read noise (e.g. from Table 2) in electrons, and the gain is the gain in electrons/DN (e.g. from Table 2).
[a href=\"index.php?act=findpost&pid=77402\"][{POST_SNAPBACK}][/a]

Yes, I do this math all the time.  The 20D's actual measured RAW noise is scaringly close (within 2% generally) to the formulaic predictions based on blackframe noise and photon count.  I don't know what you're trying to say by mentioning it, though.

[bjanes]

Not surprising, since he only charted areas affected mainly by photon shot noise on the 1DmkII.
[a href=\"index.php?act=findpost&pid=77410\"][{POST_SNAPBACK}][/a]

Since shot noise with the ID MII is the limiting factor over the dynamic range contained in a real world capture, why chart lower values where readout noise predominates?

As you imply below, dynamic range is defined by electronic engineers as the ratio of full well capacity to read noise, both expressed in electrons. Using Roger's values of 52,300 and 16.6 respectively gives a DR of 3151 or 11.6 f/stops, which is pretty optimistic under real world photographic conditions. The noise floor of engineers is somewhat different from that of photographers, who are more concerned with shadow noise as judged subjectively or perhaps expressed as a standard deviation of pixel values.

This chart is not for determining DR, of course; DR is the ratio of maximum signal to the lowest usable signal, or the noise floor (readout noise), in the same units.
He is comparing tiny-pixel cameras to large-pixel cameras in this chart.
[a href=\"index.php?act=findpost&pid=77410\"][{POST_SNAPBACK}][/a]

No one said the chart was for DR. The whole point was to compare a large pixel camera to one with small pixels.

I thought we were talking about 5D and 1DmkII-sized pixels to 8MP APS-sized pixels.  That was what all my comments were about, anyway. 
[a href=\"index.php?act=findpost&pid=77410\"][{POST_SNAPBACK}][/a]

Nowhere did you specify this. If you want to compare large and small pixels, the ID MII and S70 is a good comparison. The analysis shows that the small pixels camera is at quite a disadvantage at ISO 100. The differences are even more pronounced at high ISOs.

Comparing the 1D MII pixels to those of the 8MP APS Canons at ISO 100 is somewhat anomalous, since the former does not make full use of its full well at ISO 100.

[Gary Ferguson]

So is bigger better?

[bjanes]

So is bigger better?
[a href=\"index.php?act=findpost&pid=77471\"][{POST_SNAPBACK}][/a]

Yes, isn't that the conclusion that Michael made, supported by Roger Clark's findings and analysis? Improvements in quantum efficiency, full well size, and read noise will allow better results with smaller pixels, but can also be applied to large pixels.

[John Sheehy]

Since shot noise with the ID MII is the limiting factor over the dynamic range contained in a real world capture, why chart lower values where readout noise predominates?

Because that is where we're trying to reach when we want more dynamic range.

If we just want less noise in the "belly" of the tonal curve, then we need *much* larger photon capacity, and actually use it properly in one ISO.

Readout noise is what keeps the lower left part of Roger's chart from following the lines suggested by the upper right part of the chart; it is what causes the down-curve on the left, which limits DR.

And a very big point - How many people are tired of the line-banding in images?  That is the most distracting, disturbing aspect of Canon deep shadows, and it is purely blackframe readout noise in most cases.  There is another type of banding, caused by differences in amplification.  There is 1.019x as much amplification in the odd lines than the even lines (or visa versa) on my Canon 10D.  White balancing the image removes the discrepancy between the blue channel and red channel (both are completely segregated in this inequity).  The alternating green gains can only be interpreted as a real luminance difference, unless the converter treats the CFA data as four distinct colors, two of which happen to be the same.  This banding shows up mostly in clipped gradients, or images resized to 66% with nearest neighbor.  It is not a characteristic of shadow areas.  The banding that we do see in shadow areas *is* readout noise, and it is a very real limiting factor to practical DR.

There is no hope for tiny-sensel cameras, except noise reduction.  They just can't count high enough to have clean lower-midtones at low ISOs, or even lower-highlights at the higher ISOs.  Practical DR would be limited just by shot noise with those cameras; for DSLRs, read noise is the biggest pest.  When you count 50,000 electrons, you're talking about shot noise that is negligible over many of the top stops.  Read noise cuts into the benefits when you go deeper into the shadows, as anyone who has ever tried to pull up ISO 100 shadows knows.

As you imply below, dynamic range is defined by electronic engineers as the ratio of full well capacity to read noise, both expressed in electrons. Using Roger's values of 52,300 and 16.6 respectively gives a DR of 3151 or 11.6 f/stops, which is pretty optimistic under real world photographic conditions.

As I said earlier, I don't think that his 16.6 electron figure can be correct.  That would make the 1DmkII a stop better than the 5D in dynamic range, and people would have noticed that, by now.  The 5D has about 29 electrons of readout noise at ISO 100.  (1.95/3524)*52300 = 28.94.  1.95 is the standard deviation of noise in a 5D ISO 100 blackframe.  3524 is the highest RAW level on the 5D after subtracting blackpoint. 52300 is an approximate count of electrons countable at ISO 100.

Now, I just had an idea.  Noise is both positive and negative about the average luminance.  For a blackframe, that means values *literally* less than black.  I suspect that the figure of 16.6 came from doing a standard deviation of a blackframe *after* clipping everything below the blackpoint to the blackpoint.

In other words, a tranformation like this:

126 -> 128
127 -> 128
128 -> 128
129 -> 129
130 -> 130

Doing that to a 5D ISO 100 blackframe I have here drops the Std dev from 1.95 to 1.16.   (1.16/3524)*52300 =  17.2 electrons, very close to the 16.6 figure.  I am assuming the same number of photons at ISO 100 on both the 5D and 1DmkII.  They should be very close in reality.  If this is the problem, then Roger's figures and the figures on some astro-photography sites are wrong.  I know that Roger works off a linear 16-bit TIFF from ImagesPlus (I have corresponded with him to great extent on usenet), and that very well may clip the blackpoint, which would really make the blackframe data useless, except for comparison to other brands of DSLR that clip at black in the RAW data (as Nikon does).

I don't think of DR as anything but a basis of comparison.  Its calculated value is not to be taken literally.  There isn't some real concrete "floor" a certain number of stops below saturation, above which things can be distinguished, and below which they can't.  You can capture detail well below the 1:1 ratio of signal and noise, and nothing unique or sudden happens to quality at the 1:1 point; it's just a frame of reference; the S/N drops the deeper you go into the shadows.  In fact, if you start binning or averaging the pixels for a smaller image with more actual or apparent bit-depth, the noise floor moves further down.  With extreme binning (and with stacked images, as well), the random noise mostly disappears, but the line-banding noise doesn't.  This means that the banding is what is limiting the practical DR of binning, and banding is a noise that is almost *absent* from the std dev of the blackframe.  Removing banding from a blackframe only lowers the Std dev by a few percent.  The real limit of binning is one that isn't even adequately accounted for in the full-res "noise floor".

[bjanes]

Because that is where we're trying to reach when we want more dynamic range.

If we just want less noise in the "belly" of the tonal curve, then we need *much* larger photon capacity, and actually use it properly in one ISO.

Readout noise is what keeps the lower left part of Roger's chart from following the lines suggested by the upper right part of the chart; it is what causes the down-curve on the left, which limits DR.

And a very big point - How many people are tired of the line-banding in images?  That is the most distracting, disturbing aspect of Canon deep shadows, and it is purely blackframe readout noise in most cases.  There is another type of banding, caused by differences in amplification.  There is 1.019x as much amplification in the odd lines than the even lines (or visa versa) on my Canon 10D.  White balancing the image removes the discrepancy between the blue channel and red channel (both are completely segregated in this inequity).  The alternating green gains can only be interpreted as a real luminance difference, unless the converter treats the CFA data as four distinct colors, two of which happen to be the same.  This banding shows up mostly in clipped gradients, or images resized to 66% with nearest neighbor.  It is not a characteristic of shadow areas.  The banding that we do see in shadow areas *is* readout noise, and it is a very real limiting factor to practical DR.

There is no hope for tiny-sensel cameras, except noise reduction.  They just can't count high enough to have clean lower-midtones at low ISOs, or even lower-highlights at the higher ISOs.  Practical DR would be limited just by shot noise with those cameras; for DSLRs, read noise is the biggest pest.  When you count 50,000 electrons, you're talking about shot noise that is negligible over many of the top stops.  Read noise cuts into the benefits when you go deeper into the shadows, as anyone who has ever tried to pull up ISO 100 shadows knows.
As I said earlier, I don't think that his 16.6 electron figure can be correct.  That would make the 1DmkII a stop better than the 5D in dynamic range, and people would have noticed that, by now.  The 5D has about 29 electrons of readout noise at ISO 100.  (1.95/3524)*52300 = 28.94.  1.95 is the standard deviation of noise in a 5D ISO 100 blackframe.  3524 is the highest RAW level on the 5D after subtracting blackpoint. 52300 is an approximate count of electrons countable at ISO 100.

Now, I just had an idea.  Noise is both positive and negative about the average luminance.  For a blackframe, that means values *literally* less than black.  I suspect that the figure of 16.6 came from doing a standard deviation of a blackframe *after* clipping everything below the blackpoint to the blackpoint.

In other words, a tranformation like this:

126 -> 128
127 -> 128
128 -> 128
129 -> 129
130 -> 130

Doing that to a 5D ISO 100 blackframe I have here drops the Std dev from 1.95 to 1.16.   (1.16/3524)*52300 =  17.2 electrons, very close to the 16.6 figure.  I am assuming the same number of photons at ISO 100 on both the 5D and 1DmkII.  They should be very close in reality.  If this is the problem, then Roger's figures and the figures on some astro-photography sites are wrong.  I know that Roger works off a linear 16-bit TIFF from ImagesPlus (I have corresponded with him to great extent on usenet), and that very well may clip the blackpoint, which would really make the blackframe data useless, except for comparison to other brands of DSLR that clip at black in the RAW data (as Nikon does).

I don't think of DR as anything but a basis of comparison.  Its calculated value is not to be taken literally.  There isn't some real concrete "floor" a certain number of stops below saturation, above which things can be distinguished, and below which they can't.  You can capture detail well below the 1:1 ratio of signal and noise, and nothing unique or sudden happens to quality at the 1:1 point; it's just a frame of reference; the S/N drops the deeper you go into the shadows.  In fact, if you start binning or averaging the pixels for a smaller image with more actual or apparent bit-depth, the noise floor moves further down.  With extreme binning (and with stacked images, as well), the random noise mostly disappears, but the line-banding noise doesn't.  This means that the banding is what is limiting the practical DR of binning, and banding is a noise that is almost *absent* from the std dev of the blackframe.  Removing banding from a blackframe only lowers the Std dev by a few percent.  The real limit of binning is one that isn't even adequately accounted for in the full-res "noise floor".
[a href=\"index.php?act=findpost&pid=77501\"][{POST_SNAPBACK}][/a]

Thanks, John, for a very informative post!

[John Sheehy]

The noise floor of engineers is somewhat different from that of photographers, who are more concerned with shadow noise as judged subjectively or perhaps expressed as a standard deviation of pixel values.

Shadow noise is not just shot noise, though.  Even at a level where the readout noise component is 25% as strong as the shot noise component, the difference is visible.  Again, Roger's figures are very kind to the 1DmkII.  It actually has more like 29 electrons worth of readout noise, and his graphs fail to reflect it.  With the real figures, you start to see mostly readout noise only about a stop below middle grey; for something like the blue channel of WB'ed incandescent light, that would be one stop *above* middle grey.  I don't know where you get this idea that read noise is small in the tones of a normal-DR image.

Here is what the chart looks like with the proper read noise level; ISO 100 for the 1DmkII and 5D, with shot noise only, read noise only, and the total (square root of the sum of their squares):



This is for the green channel, which is the most sensitive.  For WB of daylight, the whole thing shifts one stop or tick to the left for the red channel.  For the blue channel of incandescent, it shifts 2 stops to the left.  It's not so deep in the shadows, before read noise makes a big difference.

Comparing the 1D MII pixels to those of the 8MP APS Canons at ISO 100 is somewhat anomalous, since the former does not make full use of its full well at ISO 100.
[a href=\"index.php?act=findpost&pid=77445\"][{POST_SNAPBACK}][/a]

You seem to say that as if everybody knows that the biggest-pixel canons don't have a large noise advantage at the 100% pixel level, over the APS-sized Canon sensor cameras with somewhat smaller pixels.  It is worth mentioning, and that was what I was discussing.  Nowhere did I mention any compact tiny-pixel cameras.  I was talking about Canon's inability to improve much on the pixel-level noise with slightly bigger pixels in their top-end DSLRs, over their more consumer-oriented DSLRs.

This is useful information, let's say, for someone who is looking for a new camera to shoot small wildlife or things at a distance - the 5D and 1DmkII are considered more professional cameras than the 20D or 30D, so it might seem logical to someone that they will get sharper, less noisey images with the pro cameras, but that is only true of the 5D, *if* you use a suitably longer lens, with equal or better resolution as a function of frame size.  If, however, you are using the same lens (as when the FOV is overkill on both cameras for the small subjects), the 20D and 30D will capture more subject detail, and the noise will be finer, and not significantly bolder at the 100% pixel view.  Of course, the larger-format cameras do have better looking viewfinders, and the 1 series have excellent burst speed, and better weatherproofing.

[John Sheehy]

Thanks, John, for a very informative post!
[a href=\"index.php?act=findpost&pid=77509\"][{POST_SNAPBACK}][/a]

Thanks for giving me cause to discover/remember the possible black clipping issue.  I think lots of people are going by black-clipped blackframe noise to get their figures, but they are too conservative.  The recent Canons all try to average 128 for black; in an ISO 100 blackframe, there are as many pixel values below black as there are above black.  They are mostly between 124 and 132, predominantly 127, 128, and 129.  The standard deviation is calculated by taking the square root of the mean of all the differences from 128, squared, since 128 is the mean pixel value.  If all values below 128 become 128 through clipping, then the mean becomes about 128.8, and more than half of the values are now 128, very close to 128.8, and all the values higher than 128.8 are now closer to the mean, so the std dev drops.  Unfortunately, as soon as the photo-signal level rises above 130 or 131, almost all of the pixels are unclipped, and all the noise is there, in its fullness, but under-estimated by anyone who measured blackframe noise from black-clipped data.

[bjanes]

Anyone with a 1DmkII and a PC with the program IRIS can verify the blackframe noise.  Just shoot at ISO 100, a very short exposure, lens cap on, and load the RAW into IRIS with the "Load a RAW file" menu item. 
[a href=\"index.php?act=findpost&pid=77410\"][{POST_SNAPBACK}][/a]

John,

I don't have the Canon but rather a Nikon D200. Previously I applied some of Roger's methodology for sensor noise and full well capacity using DCRaw to convert the images and a freeware program called ImageJ to do the numerical analysis, since I did not have the ImagesPlus program that Roger uses. The problem was that with all the DCRaw switches I was never sure if I was getting the unaltered raw data from the camera.

Roger recently published some data on the D200 which appears a bit suspect to me and I would like to try to reproduce the data for myself.

At your suggestion I downloaded IRIS and it seems well suited to the task. Is that what you are using and recommend for this type of work?

Yes, I do this math all the time.  The 20D's actual measured RAW noise is scaringly close (within 2% generally) to the formulaic predictions based on blackframe noise and photon count.  I don't know what you're trying to say by mentioning it, though.
[a href=\"index.php?act=findpost&pid=77410\"][{POST_SNAPBACK}][/a]

Yes, I do see that you know what you are talking about. Sorry, but I am just beginning to learn about some of these matters.  


Thanks for your help.

Bill

[John Sheehy]

I don't have the Canon but rather a Nikon D200. Previously I applied some of Roger's methodology for sensor noise and full well capacity using DCRaw to convert the images and a freeware program called ImageJ to do the numerical analysis, since I did not have the ImagesPlus program that Roger uses. The problem was that with all the DCRaw switches I was never sure if I was getting the unaltered raw data from the camera.

Since we first spoke about DCRAW on the RG forum about a year ago, DCRAW has grown a new switch, "-D" (capital D) which outputs the RAW data verbatim.  However, if you save in photoshop format, photoshop mangles the data upon loading into its "15-bit plus 1" number space.  You can save out in PPM format and I believe maybe even TIFF now, but the problem remains - if you don't load them into a  program with a real 16-bit number-space, they will get posterized, as DCRAW outputs the numbers 0 through 4095.

That's why I just work in IRIS as much as possible; it imports the RAW data literally, and lets you multiply it, etc, before outputting it.  The command "split_cfa a b c d" will separate the image into sub-images based on the CFA filters, where

cd
ab

represent the 4 pixels from the lower left corner.  With most (or all) Canons, that would be

rg
gb

and so I use "split_cfa g2 b r g1".  For the nikons, the greens are where the red and blue are in the Canons, and the red and blue vary with camera model, IIRC.  

For a CFA-filtered real exposure, the colors must be separated for any kind of noise analysis, because the difference in response of the color filters has a standard deviation in and of itself (and focused detail as well).

If you want to see the noise at various RAW levels, shoot a color checker completely out of focus with even lighting on it.  Then, separate the channels, and use the mouse to make a square within the squares, right-click within the square and select "stat" from the menu; the statistics for that area will then be displayed.  Between the three color channels and all the squares, you will have a large sample of intensities.

Roger recently published some data on the D200 which appears a bit suspect to me and I would like to try to reproduce the data for myself.

I last corresponded with him on usenet in June about something like that - he looked again and said that his software did not work as well with Nikon RAWs as with Canon RAWs in some regard, but I don't remember exactly what it was about.

At your suggestion I downloaded IRIS and it seems well suited to the task. Is that what you are using and recommend for this type of work?

So far, it is the most useful I have found for looking at RAW data.  I just downloaded MaximDL last night (commercial demo), which has some similar functionality, and user-created plugin capability.  Seems a lot slower than IRIS, but has a more modern interface.

[a href=\"index.php?act=findpost&pid=77529\"][{POST_SNAPBACK}][/a]

It seems that some basis of comparison is needed to equalize black-clipped and non-black-clipped data, since comparisons can be made between cameras whose RAW data is either clipped or unclipped.  Hopefully, a simple relationship exists so that clipped and unclipped blackframe noise values can be translated from one to another.  I suppose that it is possible with noise that has the same basic histogram shape, as long as the clipping occurs at the true mean.

[John Sheehy]

It seems that some basis of comparison is needed to equalize black-clipped and non-black-clipped data, since comparisons can be made between cameras whose RAW data is either clipped or unclipped.  Hopefully, a simple relationship exists so that clipped and unclipped blackframe noise values can be translated from one to another.  I suppose that it is possible with noise that has the same basic histogram shape, as long as the clipping occurs at the true mean.
[a href=\"index.php?act=findpost&pid=77536\"][{POST_SNAPBACK}][/a]

I did some clipping of a few different cases of bell-curve-shaped noise histograms, and in all cases the std dev was reduced to 58% - 59% of the full curve.  That accounts very closely for the very conservative noise figures in Roger's page and in some of the data on the IRIS website.

It might seem at first glance that the clipping is actually beneficial; after all, anything darker than black is clearly wrong and should be black.  The problem is, you also clip small signals down there, in the range where the noise is clipped.  I will try to do some extreme pushing experiments and see what the differences are between black-clipping before and after demosaicing, binning, or downsampling.  I do already know that black-clipped blackframe data gets brighter (no rendered black) from downsampling, demosaicing, and binning.

[John Sheehy]

I did some clipping of a few different cases of bell-curve-shaped noise histograms, and in all cases the std dev was reduced to 58% - 59% of the full curve.  That accounts very closely for the very conservative noise figures in Roger's page and in some of the data on the IRIS website.[a href=\"index.php?act=findpost&pid=77793\"][{POST_SNAPBACK}][/a]

I've looked at Roger's page on the 1DmkII, and he has a chart of the blackframe noise for all ISOs, and the figure for 1600 seems to be realistic, but they get to the "too good" level at ISO100, so black-clipping can't explain the values for all ISOs.  I wish I had a set of blackframes at all ISOs from one, so I could see if the low readout noise is for real at ISO 100.  Now, the IRIS website claims the same level for the 5D as Roger claims for the 1DmkII, and that clearly isn't true.