canon 1d mark III

[John Sheehy]

I don't think we should assume that 14-bit processing is in itself is a huge improvement, but rather it is indicative of improved shadow noise and dynamic range which requires the additional 2 bits for the capture of which.
[a href=\"index.php?act=findpost&pid=102895\"][{POST_SNAPBACK}][/a]

14 bits isn't going to help a lot, unless read noise is reduced.  Based on the samples, it doesn't really seem like there is any reduced noise in the MKIII.  Hopefully, the production models will do better than the cameras that made the samples.

[phila]

Video of the MkIII & 580 II:

www.photo-i.co.uk/News/Feb_07/Canon_spring.html

[phila]

14 bits isn't going to help a lot, unless read noise is reduced.  Based on the samples, it doesn't really seem like there is any reduced noise in the MKIII.  Hopefully, the production models will do better than the cameras that made the samples.
[a href=\"index.php?act=findpost&pid=102899\"][{POST_SNAPBACK}][/a]

From IR:

"...Now that we've had some hands-on time with a late-rev prototype of the Canon EOS-1D Mark III, we're more impressed than ever. While the sample we've been working with is only a prototype, it's image quality and high-ISO noise levels are simply extraordinary. To our eyes, it looks like Canon has managed almost a full f-stop of noise improvement, the Mark III's ISO 6400 shots are only a little noisier than ISO 3200 ones from the Mark II N..."

Lots of sample images:

www.imaging-resource.com/PRODS/E1DMK3/E1DMK3A5.HTM

www.imaging-resource.com/PRODS/E1DMK3/E1DMK3A7.HTM

[John Sheehy]

From IR:

"...Now that we've had some hands-on time with a late-rev prototype of the Canon EOS-1D Mark III, we're more impressed than ever. While the sample we've been working with is only a prototype, it's image quality and high-ISO noise levels are simply extraordinary. To our eyes, it looks like Canon has managed almost a full f-stop of noise improvement, the Mark III's ISO 6400 shots are only a little noisier than ISO 3200 ones from the Mark II N..."

Lots of sample images:
[a href=\"index.php?act=findpost&pid=102913\"][{POST_SNAPBACK}][/a]

The JPEGs at high ISO look less noisy, but they also look a little softer, so I don't know how much is due to NR in the JPEGs.  I really need to see RAW files to come to a conclusion.  The low-ISO images have very noisy shadows, IMO; I don't know how JPEG compression is affecting them.  Improving high-ISO doesn't necessarily help on the low end, and Canon seems to be concentrated on the high end.

[BJL]

Well, do you have any information on what the fill factor of the Canon CMOS chips is? Even if the fill factor approaches 100 percent and the transistor area of the chip is markedly reduced, you still need a sufficient active pixel area to capture photons. The quantum efficiency could be doubled or tripled, but you still have to have a sufficient pixel size to control shot noise and maintain and dynamic range and sensitivity.  These factors are summarized by Roger Clark. Do you have this additional information?

Bill
[a href=\"index.php?act=findpost&pid=102697\"][{POST_SNAPBACK}][/a]

Canon does not disclose such details, but companies like Sony, Kodak and Dalsa, who sell sensors directly, tell us more about their sensors. The new Sony sensor has photosites of about 12% the area of those in the 1DMkIII (2.5 vs 7.2 microns) and the well occupies about 30% of that area (effective fill factor is raised to about 50% by microlenses). So the transistors in that Sony CMOS sensor occupy an area of at most about 8% of the total area of a 1DMkIII photosite. And that is with "largish" 180nm process: the space occupied by transistors could potentially be reduced by a factor of nine by changing to the current state of the art 60nm process.

The possibilities for feature size reduction well below current levels has to be taken into account before predicting that we are at or near the limits of minimum viable photosite size. The paper by Wandell et discuss that factor?


Also, at high ISO speeds, wells are not being filled even at highlights; the sensor is effectively "underexposed". So well capacity is rather irrelevant to high ISO noise: what counts is how many photons the camera's lens can deliver to each photosite, the efficiency of the photosites at detecting those photons, and dark noise/read noise level in electrons.

Well size is relevant to maximum dynamic range possible when wells are almost filled at highlights, which is an aspect of performance at low ISO speeds.

[BJL]

... there are limits there too as shown in a recent thread comparing Kodak and Dalsa chips. If the well is too deep, edge performance suffers.
[a href=\"index.php?act=findpost&pid=102737\"][{POST_SNAPBACK}][/a]

The comparison shows the opposite when you skip the marketing hype and read the data sheets. The Kodak 39MP sensor with "narrower, deeper" wells of larger electron capacity has equally good off-perpendicular sensitivity as the Dalsa sensor with "wider, shallower" wells of smaller electron capacity.

[BJL]

According to Roger Clark, electron density per square micron of chip area has not improved much over the years
[a href=\"index.php?act=findpost&pid=102737\"][{POST_SNAPBACK}][/a]

According to actual data instead, there have been significant improvements over the last few years.

Kodak has increased the well capacity of its 6.8 micron pixel pitch FF CCD sensors 50%, from 40,000e in the sensor of the Olympus E-1 to 60,000e in recent models such as in the Leica M8 and H3D-39. That is, from 865e per square micron to 1300.

Also, the newly announced Sony CMOS sensor with 2.5 micron pitch has a well capacity of 10,000e (min), or 1600 per square micron. Quite impressive given the twin disadvantages of the lower fill factor of CMOS (30% I think Sony says, vs over 50% for the Kodak FF CCD's) and the expected lower fill factor of smaller photosites. The wells of the Sony sensor must be even "deeper" than the new, deeper than before Kodak ones.

[bjanes]

According to actual data instead, there have been significant improvements over the last few years.

Kodak has increased the well capacity of its 6.8 micron pixel pitch FF CCD sensors 50%, from 40,000e in the sensor of the Olympus E-1 to 60,000e in recent models such as in the Leica M8 and H3D-39. That is, from 865e per square micron to 1300.

Also, the newly announced Sony CMOS sensor with 2.5 micron pitch has a well capacity of 10,000e (min), or 1600 per square micron. Quite impressive given the twin disadvantages of the lower fill factor of CMOS (30% I think Sony says, vs over 50% for the Kodak FF CCD's) and the expected lower fill factor of smaller photosites. The wells of the Sony sensor must be even "deeper" than the new, deeper than before Kodak ones.
[{POST_SNAPBACK}][/a]

Your figures are similar to what Roger reports: "For current technology of CCD and CMOS sensors, the full well capacities run about 800 to 1600 electrons per square micron. These values haven't changed much on over twenty years of sensor development."

Of course, you have to take fill factor into account and not simply consider pixel pitch. Fill factor for CCDs is typically 80-90% ([a href=\"http://www.dalsa.com/pi/products/DSC.asp]Dalsa[/url]) and is much lower for CMOS, around 30% for many designs, but improving as the transistor size is reduced by improved manufacturing techniques. We are not talking about an order of magnitude improvement here.  

Bill

[John Sheehy]

Well size is relevant to maximum dynamic range possible when wells are almost filled at highlights, which is an aspect of performance at low ISO speeds.
[a href=\"index.php?act=findpost&pid=103034\"][{POST_SNAPBACK}][/a]

And then, the lowest ISO needs to be amplified in such a way that the read noise is low, relative to max signal.  If the RAW data only goes up to 2700, and the read noise is basically the same as ISO 100 in ADUs, then you're never going to see any improvement in DR; just better quality highlights and midtones.

[Ray]

And then, the lowest ISO needs to be amplified in such a way that the read noise is low, relative to max signal.  If the RAW data only goes up to 2700, and the read noise is basically the same as ISO 100 in ADUs, then you're never going to see any improvement in DR; just better quality highlights and midtones.
[a href=\"index.php?act=findpost&pid=103098\"][{POST_SNAPBACK}][/a]

I think all of us who have taken the trouble to compare images at ISO 1600 from our 20D, 5D or 1D2 with the same scene underexposed by 4 stops at ISO 100, are impressed with the huge improvement in noise in the ISO 1600 shots, which stretches right across the tonal range but is particularly significant in the deep shadow, lower mid-tones and mid-tones. Owners of the new 1D3 should be able to compare ISO 3200 with a 5 stop underexposed ISO 100 shot and expect to see an even greater improvement, unless Canon have also improved shadow noise at ISO 100, in which case the degree of noise improvement might just be the same as that of the 1D2, just improved equally at all ISOs. I guess that remains to be seen.

However, I have to admit, as much as I am impressed with the results, I haven't much of a clue as to what processes are employed to achieve these outstanding results. Perhaps some of you clever chaps can enlighten me.

I have some vague understanding that the voltages generated at each photosite are amplified in the analog domain before becoming digitised. Such amplification of the signal allows read noise to be lower than it otherwise would be with an unamplfied (or less amplified) signal. (Don't know why this should be the case, though.)

I also have some vague understanding that the initial analog signal needs to be amplified because, in relation to full-well capacity, it's a much weaker signal than it could be.

What I don't understand at at all is, having established the technique of reducing noise through analog amplification (and no doubt through another 100 patented electronic tricks) why should well capacity limit the degree of amplification?

I have, for example, a shot at ISO 100. Full-well capacity is, say 50,000e which represents the highlights with ETTR exposure. I want to achieve the same level of noise in the shadows as I get with ISO 1600 whilst preserving the highlights and dynamic range that one expects at base ISO.

One could argue that boosting the analog signal in this situation would 'blow out' the highlights. Why should it? One photosite generates its maximum 50,000e signal and when amplified 4x becomes a 200,000e signal. Another photosite generates a 40,000e signal which, when amplified 4x becomes a 160,000e signal. The relationship and relativities are preserved.

Where's the problem? Is it possible that the interconnects cannot sustain such high voltages, or that excessive heat would degrade the results?

Can someone do me a favour and clarify this issue?

[John Sheehy]

One could argue that boosting the analog signal in this situation would 'blow out' the highlights. Why should it? One photosite generates its maximum 50,000e signal and when amplified 4x becomes a 200,000e signal. Another photosite generates a 40,000e signal which, when amplified 4x becomes a 160,000e signal. The relationship and relativities are preserved.

Where's the problem? Is it possible that the interconnects cannot sustain such high voltages, or that excessive heat would degrade the results?

[a href=\"index.php?act=findpost&pid=103141\"][{POST_SNAPBACK}][/a]

There has to be a target voltage range for the ADC.  Perhaps most of the read noise at low ISOs happens after the amplification at the photosites; before or in the ADC.  IOW, perhaps part of the read noise has nothing whatsoever to do with amplification, and composes most of the total read noise at ISO 100, and only a fraction of it at ISO 1600.

[bjanes]

One could argue that boosting the analog signal in this situation would 'blow out' the highlights. Why should it? One photosite generates its maximum 50,000e signal and when amplified 4x becomes a 200,000e signal. Another photosite generates a 40,000e signal which, when amplified 4x becomes a 160,000e signal. The relationship and relativities are preserved.

Where's the problem? Is it possible that the interconnects cannot sustain such high voltages, or that excessive heat would degrade the results?

Can someone do me a favour and clarify this issue?
[{POST_SNAPBACK}][/a]

The pre-amplifiers in the chip amplify voltage, not charge. 50,000 electrons in a photosite will produce a voltage according to the capacitance of the site, according to the formula V = Q / C, where V is voltage, Q the charge (1 coulomb = 6.24 × 10^18 electrons), and C is the capacitance. In a CMOS chip, this voltage is amplified by the transistors in the pixel, and the output of the pixel is voltage. This voltage is further amplified before being presented to the A/D converter as shown in this [a href=\"http://www.dalsa.com/shared/content/Photonics_Spectra_CCDvsCMOS_Litwiller.pdf]diagram[/url], figure 2.

At base ISO, the amplifier gain is usually set so that the full range of the ADC matches the single-pixel linear full well capacity of the chip. If one is shooting at 1/16 of base ISO (1600 for a camera with base ISO of 100), the well of the sensor fills only to 1/16 of full well. The amplifier gain is then increased 16 fold so that the output of the A/D converter will still be full scale. The number of electrons is not amplified, but rather the voltage. Dynamic range is reduced and noise increases according to photon counting statistics and read noise.

Most cameras allow a bit of headroom so that the ADU output for full well will be somewhat less than the full scale of the sensor (4095 for a 12 bit device) at base ISO. John is implying that the full well ADU of some Canons is only about 2700. In this case effective 11 bits (log[2700, base2]).

With my Nikon D200 with an exposure such that an 18% gray card gives a pixel value of about 118 with a gamma 2.2 space, the raw pixel value for the gray is 671 ADU and white paper gives 3558. Overflow occurs at 3986.

I hope this helps.

Bill

[Ray]

At base ISO, the amplifier gain is usually set so that the full range of the ADC matches the single-pixel linear full well capacity of the chip.

Well, thanks Bill for pointing out that the 'photon to electron' charge has to be converted to a voltage before analog amplification. I missed that point, but it doesn't really change the question, 'Why does the amplifier gain have to be set so the full range of the ADC matches full well capacity?'

If there's some noise advantage in amplifying the derived voltage from a charge just 1/16th and less, of full well capacity (as in an ISO 1600 shot), then why not use the same procedure to amplify the voltage derived from a full well signal. If necessary, such a signal could be brought back down to a lower level before recording to memory and hopefully leave some noise behind in the process.

Are the reasons just the impracticality of finding room on the sensor for perhaps the bigger and heavier duty components that might be required?

[bjanes]

Well, thanks Bill for pointing out that the 'photon to electron' charge has to be converted to a voltage before analog amplification. I missed that point, but it doesn't really change the question, 'Why does the amplifier gain have to be set so the full range of the ADC matches full well capacity?'

[{POST_SNAPBACK}][/a]

Most current ADCs in common cameras have 12 bits resolution, but this goes up to 16 bits in high end units. As you know the new Canon uses 14 bits. If the maximal voltage is only half scale on the ADC, you have lost half of its resolution. When operating at higher than base ISO, the sensor wells are only partially filled, but still you want to use the full scale of the ADC. Thus, additional amplification is used. These considerations are explained further on [a href=\"http://en.wikipedia.org/wiki/Analog-to-digital_converter]Wikipedia[/url]

If you used a 16 bit ADC for a sensor with a full well of 50,000 electrons, each electron could be counted directly (gain = 1 ADU per electron), further amplification omitted, and you could adjust ISO in the raw converter. According to Roger Clark, it makes no sense to increase ISO on the camera above unity gain (ISO 1600 in the 5D, 800 on the D200, 1 ADU = 1 electron). John Sheehy feels better with some degree of oversampling. You can increase the ISO in the raw converter with no adverse effects in this case, just as we do with white balance. Since you like to think out of the box, another interesting idea would be to use an ADC with a log output. In this case information density would be spread evenly over the scale, and half the information would not be in the brightest f/stop as with current linear ADCs.

Bill

[Ray]

If you used a 16 bit ADC for a sensor with a full well of 50,000 electrons, each electron could be counted directly (gain = 1 ADU per electron),

Are you saying the problem is a matter of bit depth? If we want to start amplifying ISO 100 signals, then we might need perhaps 32 bit (per color) processing?

When operating at higher than base ISO, the sensor wells are only partially filled, but still you want to use the full scale of the ADC. Thus, additional amplification is used. These considerations are explained further on Wikipedia

This Wikipedia article seems to be mainly about audio ADC converters. I'm having difficulty in seeing the relevance.

The noise advantage in Canon high ISO images seems to be due to analog amplifier gain, which is denied the ISO 100 signal.

If you want to use audio analogies, I'm reminded of the Dolby B NR system on my first cassette tape recorder many years ago. The signal prior to recording was boosted and the dynamic range compressed (to fit the range on the tape). The recorded signal was thus greater in relation to system noise. On playback, both the signal and 'tape hiss' were reduced in amplitude and the dynamic range of the original signal expanded.

However, I can't relate such a process to the recording of light signals. We have no method of amplifying light waves on this microscopic scale. I don't see how audio analogies are relevant.

[BJL]

Your figures are similar to what Roger reports: "For current technology of CCD and CMOS sensors, the full well capacities run about 800 to 1600 electrons per square micron.
[a href=\"index.php?act=findpost&pid=103064\"][{POST_SNAPBACK}][/a]

Can you give a lin to what Roger Clarke says? Is Clarke saying that "electron densities" of 1600e/sqmicron was available 20 years ago? That does not fit with what I see for even high end Full Frame type CCD's from just three years ago, where Kodak had about 800 with 6.8 micron pixels and 1100 with 9 micron pixels in MF sensors (size does help, as the lateral overflow drains of larger pixels take up a smaller fraction of total photosite area.) If sensors offered 1600 up to 20 years ago, they were probably special scientific ones without lateral overflow drains: absence of LOD's is good for very high DR scientific usage, but bad for blooming.

To repeat, what I see is not close to absence of progrs over 20 years, it is
- a 50% increase in well capacity in the same FF CCD pixel size over the last several years.
- Three years ago, 1100e/sqmicron in 9 micron FF CCD pixels and only 800 in 6.8 micron FF CCD pixels, implying considerably less that 800 in digicam size pixels, especially CMOS ones with their inherently lower true fill factor.
- Today, 1600 in 2.5 micron CMOS digicam pixels.

Those changes over just a few years, suggests a substantial improvement, at least in how small photosites can be and still maintain high levels of "electrons per unit area", probably a useful measure of highlight headroom. This seems relevant to the pessimistic claims that photosites (even DSLR photosites) have got about as small as "physical limits" allow them to be and still function well.


P. S. These numbers also suggest that reduced feature size is greatly reducing the disadvantage of CMOS and interline CCD sensors relative to FF CCDs in fill factor and electron well capacity for given pixel size, which might mark the beginning of the end of FF CCD. Maybe that is why Kodak has for the first time used interline CCD instead of FF CCD in a sensor for astro-photography and FourThirds cameras, the new KAI-10100.

[BJL]

However, I can't relate such a process to the recording of light signals. We have no method of amplifying light waves on this microscopic scale.
[a href=\"index.php?act=findpost&pid=103223\"][{POST_SNAPBACK}][/a]

You amplify the signal from the photosite, just as Dolby, tape bias and RIAA bias (for vinyl) amplify high frequency parts of the signal from the microphone. In each case, it is for protection against noise arising at later stages, not for reducing "shot noise" in the original detected signal.

Maybe the future of CMOS is amplifying on site so much that the voltage per electron is comfortably above the levels of subsequent noise sources, and than having an A/D convertor that can handle a high maximum input voltage and many bits: 16 or more. Well capacities of up to 65,536 electrons (above that of all current sensors except 9 micron pixel MF sensors) could in principle be counted "exactly" in 16 bits, allowing a fixed charge to voltage conversion ratio regardless of exposure index, so effectively a fixed sensitivity in the analogue domain. ISO speed would be applied only in conversion from raw.

[bjanes]

Are you saying the problem is a matter of bit depth? If we want to start amplifying ISO 100 signals, then we might need perhaps 32 bit (per color) processing?
This Wikipedia article seems to be mainly about audio ADC converters. I'm having difficulty in seeing the relevance.

However, I can't relate such a process to the recording of light signals. We have no method of amplifying light waves on this microscopic scale. I don't see how audio analogies are relevant.
[a href=\"index.php?act=findpost&pid=103223\"][{POST_SNAPBACK}][/a]

The Wikipedia article does not even mention audio until well into the discussion. Digital signal processing is similar for both audio and video. Norman Koren (author of Imatest) got his start in audio engineering and has applied his expertise to digital photography. Much of the seminal work in signal processing got its start in Bell Laboratories and was related to sound.

The role of amplification prior to the ADC stage is to match the output voltage to the voltage range of the ADC. Bit depth has to do with resolution, the number of steps at which the highest voltage is counted. Bruce Fraser used an analogy to a staircase: dynamic range is the height of the staircase, whereas bit depth is the height of each step, that is the resolution.

The original Dolby system did not modulate sound directly, but the electrical analogue of the sound waves. I think the process is now done digitally. The same thing can be done with light. One can transform the spatial domain into the frequency domain.

Bill

[John Sheehy]

If you used a 16 bit ADC for a sensor with a full well of 50,000 electrons, each electron could be counted directly (gain = 1 ADU per electron), further amplification omitted, and you could adjust ISO in the raw converter.

In order for that to work optimally, there would have to be virtually no read noise.  Read noise is often measured in units of electrons, but it has absolutely nothing to do with electron charges.  If you could compare the actual electrons in the well, and what you get after readout, the discrepancies would not be in whole, integer electron units.  They would be in all kinds of whole-plus-fractional units of "electrons".  So, counting electrons is not really possible, in the purest sense, with any kind of read noise.  Probably about 0.3 electrons of read noise is what it would take to clearly discriminate numbers of electrons (given enough ADUs to discriminate).  And then, you'd want slightly more than 1 ADU per electron - that gives an unevenly-gapped histogram, but at least the levels are all distinct, and an algorithm could clean up the histogram by using exactly 50,001 levels (0 to 50,000).  When read noise is significant, you need even more ADUs to keep the digitized count as close as possible to the real electron count; truncating noise never decreases noise; truncating noise increases noise.  Of course, these are subtle gains in extreme shadows.

A camera that literally counted exact numbers of photons would be a dream come true; all that would be left is shot noise, and shot noise is not really a noise except that we don't want to see it, but it is a physical reality of capturing light in bins; it is not a technological problem, except for the efficiency of collection.

According to Roger Clark, it makes no sense to increase ISO on the camera above unity gain (ISO 1600 in the 5D, 800 on the D200, 1 ADU = 1 electron). John Sheehy feels better with some degree of oversampling.

Not only do I; so does the noise, especially the horizontal line banding in Canon RAW data (other ISOs are normalized for ISO 100; the total noise is divided by ten):



These trends certainly suggest that a gain-based 3200 could easily have reduced noise, and decreases horizontal banding noise, especially.

You can increase the ISO in the raw converter with no adverse effects in this case, just as we do with white balance. Since you like to think out of the box, another interesting idea would be to use an ADC with a log output. In this case information density would be spread evenly over the scale, and half the information would not be in the brightest f/stop as with current linear ADCs.
[a href=\"index.php?act=findpost&pid=103214\"][{POST_SNAPBACK}][/a]

I don't think it's really an issue of how much "information" is in the stop; the real issue is where your signal is, relative to the noises.  If you use the full top stop, read noise, relative to signal is only half what it would be if you exposed 1 stop lower, and shot noise (relative to signal), is only 71%.  The noises are far more limiting, in practice, than "numbers of values" are.

As far as log output is concerned (I think you probably really mean gamma-adjusted; 0 has no log), the real issue is the read noise, and having a gamma-adjusted output from an ADC is not going to reduce the signal-to-read noise ratios, and the diodes or transistors used would probably add more noise of their own.

[John Sheehy]

Bruce Fraser used an analogy to a staircase: dynamic range is the height of the staircase, whereas bit depth is the height of each step, that is the resolution.[a href=\"index.php?act=findpost&pid=103242\"][{POST_SNAPBACK}][/a]

That's not a particularly fitting analogy.  The "height" is meaningless without a frame of reference.  DR has various standards as to what is an acceptable S:N ratio, but basically, the idea is based upong the ratio of the highest recordable signal, to the lowest usable signal ("usable" open to interpretations and standards).