Noise is sensors

[watsok]

I have background in the design and building of CCD sensors. I see a lot of discussion about noise vs image quality vs exposure and almost anything else. I thought a simple description of just where the noise comes from and what controls do you have over it when taking an image might be of interest.

Noise comes from several places in CCD design. Thermal noise, shifting noise, amplifier noise, and Analog to Digital converter noise. The only noise in a image that a photographer has control over is Thermal noise. I will explain Thermal noise first and maybe touch on the others as most here have no interest.

The way any solid state sensor works is that they convert photons into electrons. These electrons are accumulated and stored as a “charge” or “signal”. This charge is then “read” by an analog to digital converter and a digital output is generated and stored. If you have a sensor with 8 million pixels you do not have 8 million analog to digital converters. Most likely there is a much smaller number, under a 1000, maybe even under 100 that do all the conversions.

That being the case, the “charge “ needs to be transported to the analog to digital converter before it becomes a digital signal of either red , green or blue value. This transportation is also called shifting because the charge is shifted to one edge or the other of the sensor to be read.

During all the time the exposure is being made and during transport to the analog to digital converter there is essentially a light leak occurring inside your camera. This light is infrared, thermal noise, from the body of the sensor itself. Infrared energy also is generating photons. These photons are added to the original charge until it reaches the converter.

Thinking about this you can see that since we can not eliminate this noise it is best to gather as many photons that we care about as fast as possible and get them to the Analog to Digital converter as fast as we can. By having a large charge ( large number of photons) the noise is almost not noticeable and getting it converted into a digital signal rapidly means even less noise is collected.

From a photographer who wants as little noise in their images as possible, it is best to have as short as possible exposure, with a f stop that fills the sensors pixels to capacity without saturating the pixels. You have little control how fast the charge is moved from the sensor to the analog to digital converter.

You should now be able to understand why dark areas have more noise than light areas. Their signal ( charge ) to noise ratio is a lot lower than in the highlights. The noise component is a big part of a dark area in an image because so few photons were collected that we have relatively a lot of thermal (an other) noise included in the analog to digital conversion.

The other sources of noise that are uncontrollable by the photographer are amplifier noise, noise from shifting the charges out, in-accuracy of the analog to digital converter, and generally on chip electronic noise generated just by all the activity occurring to move the image off chip.

[eviltech]

very interesting, thanks for the knowledge.

[wolfden]

learn something new everyday :-)

[Scottes]

From a photographer who wants as little noise in their images as possible, it is best to have as short as possible exposure, with a f stop that fills the sensors pixels to capacity without saturating the pixels. You have little control how fast the charge is moved from the sensor to the analog to digital converter.
This would seem to imply that it would be better to take a shot at f/2.8 ISO 6400 for 1/2000 rather than an equivalent shot of f/2.8, ISO 100, 1/30?


I'd have to take a couple test shots, but I'm willing to bet that the ISO 6400 shot will have more noise. I would think that the reduction of SNR would have more impact than the reduction of time.

[Josh_30]

Does this still apply to CMOS sensors as well? I would assume so, but I am a bit confused.

I thought that in CMOS sensors, each pixel has a charge-to-voltage converter on-chip, such that the sensor itself outputs digital information with less need for a converter after the fact when compared to CCD's? This seems as if it would reduce the light-leak you speak of as well.

So I guess my question is... does shifting and thermal noise still apply to CMOS sensors like it does to CCD's? If so, is it to a lesser degree?

:confused: I'm not an engineer... just going on what I've read.

[Ed The Shed]

So in essence shoot to the right of the histogram to avoid the noise but avoid clipping ????

Then bring the exposure back down in the raw convert ???

Ed

[Scottes]

So in essence shoot to the right of the histogram to avoid the noise but avoid clipping ????

Then bring the exposure back down in the raw convert ???
That's what I follow, but that goes against a couple statements made by watsok. "Shooting to the right" is not "as short as possible exposure" since this generally means that you will over-expose which means taking a shot with a longer shutter speed.


I understand that watsok is trying to explain things, and all of his information is correct - once you weed out some of the conflicting statements like "it is best to have as short as possible exposure".


In my opinion:
In order to get the best exposed shot with the least noise possible, shoot at the lowest ISO with the widest aperture in order to saturate the sensors as much as possible, and then adjust the exposure in the raw converter. Note that changing the aperture will have negative effects on the noise, since this will lengthen the exposure, but that is a price to pay to get the shot you want. Also, over-saturating sensors (clipping / blown highlights) is subjective, and you may find that you want to clip some areas in order to minimize noise in the important areas.


But I don't think that what I said is going to be clear to everyone.

[gjl711]

Very nice write up and welcome to the forum. There is one other aspect that needs to be discussed and that is the gain of the amplification, or why a ISO 6400 1/2000 exposure will look worse than a ISO 100 1/30 exposure. Maybe you can touch on that a bit as well.

[watsok]

Scotties,

I think if you consider how ISO is set in the camera, it is really a “gain” setting for the conversion amplifiers. The charge buckets in the sensor remain the same. At a high ISO setting the sensor is pre-setting itself to add more gain to the number of photons/electrons it will be processing into a color. In fact you are not saturating each pixel’s capacity to hold as many electrons as possible. With the amplifier gain set high ( high ISO setting), the amplifiers gain setting would distort a full charge into noise itself because they were presented with a charge much larger than expected.
The high gain setting/ high ISO setting essentially says that you plan on getting a lot fewer electrons to measure therefore they will amplify what they have to present a “Normal” curve you see in the histogram.

Low ISO setting tells the sensor you plan to fill the charge buckets up, set the amplifiers to low and let the noise drop out during conversion …almost. Low ISO setting, fast shutter speed, small numeric F stop should yield the lowest noise.

My reasoning for short exposure that fills up the charge wells fully as being best relates to having as little time as possible go by until that last charge of the image is read off the sensor. A short exposure of 1/2000th of a second vs 1/500th of a second is a long time from a electronics point of view ( 3 million clock cycles in your CPU). I believe most of the amplifiers and analog to digital converters are on the same chip as the sensor. My experience is with CCD type of sensors which us a bucket brigade sort of arrangement to get the charge to the amplifiers and converter. With CMOS sensor it is more of a direct addressing ( X-Y) to read each pixel. In both cases IR photons, amplifiers, and on chip processing cause noise. In all sensors there is a well in the silicon that is used to store a charge until it is read.

[Scottes]

OK. But which will produce less noise, a shot at f/2.8 ISO 6400 for 1/2000 or the equivalent shot of f/2.8, ISO 100, 1/30?

But you kinda answer this with the sentence "Low ISO setting, fast shutter speed, small numeric F stop should yield the lowest noise." However, this would imply that setting 1/8000 regardless of the meter setting would produce the lowest noise. Of course, that won't. If the meter reading recommends 1/30, then 1/8000 isn't going to record a thing.

You have some great explanations here, but your unconditionalized statements like the ones I've quoted are very, very bad things to be saying to a group that doesn't completely understand the context.

[Ed The Shed]

I think what we have is basically have a balancing problem.

At one end we have the requirement "electronically" to generate the least noise and at the other the photographers requirement to expose correctly for the image.

Unfortunately as we can see they are at opposite ends when it comes to low light shots.

Our job is to judge the exposure the best we can for the shot given the limitations of the light and how the camera operates.

Going back to my earlier comment on "Exposing to The Right", your right that any increase in signal will also cause an increase in noise as well. However the increase in the signal will be far greater than the increase in the noise. (Increasing Signal to noise ratio (a good thing)).

Hope you can follow that ????

Ed

[watsok]

"However, this would imply that setting 1/8000 regardless of the meter setting would produce the lowest noise. "
The purpose of obtaining a correctly exposed image is most important to photographers. This is what meter readings give as a guideline to accomplish this. If you could set the camera to 1/8000 sec. @ ISO of 100 and get a correct exposure this most likely would produce the lowest noise image.

"Of course, that won't. If the meter reading recommends 1/30, then 1/8000 isn't going to record a thing."

Here you are correct because there would not be enough time to collect enough photons to fill the sensors storage locations/pixels. To have a 1/8000 exposure when the metering is telling you need to be @ 1/30th of a second to fill the sensor storage pixel locations is self evident that you will get no or a very poor image. First you are collecting a lot fewer photons than @ 1/30th of a second, second the gains are set low inside the camera and are expecting to convert many more electrons to a digital number. This number is destined to be small, the image will have no dynamic range.

Another way of looking at ISO setting in you camera is dynamic range expansion/management. At the lowest ISO setting the dynamic range of the sensor is at maximum therefore there is no need for expansion/management.

As the ISO setting is increased, with fewer physical photons/electrons the camera adjusts itself to give maximum dynamic range ( your histogram) but it is delivering this range from a source that has less information. That is to say the ISO setting along with the meter reading from the camera are working to give a generally wide dynamic range from the expected light/ photons available. As you try to keep the dynamic range constant ( correct exposure) but with source data that keeps shrinking that is used to derive the range, quality is declining. You are attempting to keep the same range with less real data ( photons).

[Scottes]

As the ISO setting is increased, with fewer physical photons/electrons the camera adjusts itself to give maximum dynamic range ( your histogram) but it is delivering this range from a source that has less information.
You seem to be saying "increasing ISO decreases the signal-to-noise ratio which will result in more noise in the final image", right?

[René Damkot]

If I read right:

"Native" ISO is better then amplifying the signal by setting a higher ISO. This overrules the supposedly lesser noise at a shorter exposuretime with a higher ISO, correct?

You also seem to be saying that (for instance) 1/4000s; f/5.6 at a given ISO gives less noise then 1/500s; f/16 at the same ISO?

I find it *very* hard to believe that that would be a meaningful (or even measurable) difference. And I'd never let an insignificant difference in noise cause me to set a different aperture then I'd needed...

Edit: Also, the only EOS to have a CCD is the 1Dino. Al modern EOS DSLRs have CMOS sensors.

[watsok]

Increasing the ISO will decrease the signal to noise ratio...usually. I do not know the exact design of the sensor. At several low ISO's, perhaps, they can collect enough photons to where noise is at the single bit or two level. This would make it almost not noticeable. As the ISO is increased noise levels start to become a larger percentage of the signal. Exactly at what ISO this begins ( beyond 100 ) and how the internal processing software is handling things has some influence on what you will see. This is where general design physics might diverge from specific design techniques.

CMOS as in "CMOS sensor" is the semiconductor process that is used to build the sensor . Today, it is the most common of all processes, therefore most cost effective. Most semiconductor companies avoid making chips larger than about 25 mm on a side due to high yield losses ( poor profits ). The profits decrease close to the square of the area. With SLR image sensor being much larger they need to be built in well characterized processes with specialized equipment. In SLR cameras expensive sensors are covered by the high cost of the end product. This is rare in consumer electronics.

The downside of CMOS processes, high quality amplifiers and analog devices are more difficult to make. Things like amplifiers and analog converters for instance. As in all things it is an engineering compromise. As processes are refined one process can leapfrog another. I would bet a small sum of money that we will hear of a new extreme CCD camera, or camera using a Bipolar process that puts all the CMOS sensor cameras to shame. Sony would be the most likely source.

[troypiggo]

While I'm not up there on a technical level with some of you guys, I'd like to add a few thoughts and things I've learned since taking up my new hobby of astrophotography.

As you can imagine, taking images of a dark sky with only stars, nebulae, comets, planets etc as light sources is quite the challenge. There is no "expose to the right". In fact, if you look at the histogram there's a massive spike at the black end of the histo and not much else. Also, if shooting with a DSLR, typical settings might be ISO800 and 4 minute, 8 min, even 10 min exposures. You can imagine the noise these images have. We spend a lot of time calibrating them to get the best possible signal to noise ratio.

How we combat all this noise is to take many exposures of the same target, same settings. Maybe tens or twenties or fifties of shots of exactly the same thing. These are called "light" frames. The more the better. This is to increase the signal to noise ratio. Stacking software then takes all the images, aligns and stacks them, keeping the signal from the stars or whatever you're shooting, and smooths out all of the random pixels that are presumably noise.

On top of that, we take "darks", "flats", "flat-darks", and "bias" frames. These are to counter some of the other noise and vignetting etc that the images have. A lot of work goes into this.

"Darks" are frames taken with, say, the telescope lenscap on. No light in. You take the shot at the same ISO and shutter speed as the "light" frames, and at the same ambient temperature as the light frames. Again, you might take 10 of these. They're similar to Canon's in camera noise reduction. They are subtracted from the light frames to remove hot pixels.

"Flats" are shots taken off an even light source and their purpose is to calibrate dust spots and vignetting. Maybe not relevant for this thread. If you want to get really carried away, you'd also take a "flat-dark" which is similar to the above darks but use the same settings as what you took the flat with.

"Bias" frames are taken again with lenscap on, at same ISO but fastest possible shutter speed. These are to pick up the noise from sensor readout etc.

Won't go into the maths of it, but these calibration frames are subtracted from the stacked light frames by stacking software and you can end up with a pretty good SNR.

Would you go to all the above trouble for "normal" photography? I wouldn't. But it does give you an appreciation of where all the different noise contributions come from as a supplement to watsok's posts above.

Another comment I'll drop here is that in the astrophotography world it's CCD cameras over CMOS DSLRs that are the really high end, top notch gear. They have much greater well depths, so it takes them longer to saturate. Plus some have anti-blooming so a full well won't affect neighbouring pixels. And they have much better quantum efficiency, so when photons do hit the sensor it's much more likely to measure. They are also in the order of $3000 up to $20,000 or more for the good ones. Wish I could afford one.

[Scottes]

Nice info Troy. It does give an idea of other ways to combat all the different types of noise. What program do you use to do all that, anyway?

[troypiggo]

Deep Sky Stacker (http://deepskystacker.free.fr/english/index.html) is very popular and free. Fairly quick and simple to use. Has pretty good explanations of what I was talking about above in the FAQ (http://deepskystacker.free.fr/english/faq.htm) and another page on how to take better astro images (http://deepskystacker.free.fr/english/theory.htm).

I'm actually using Nebulosity (http://www.stark-labs.com/nebulosity.html) now. It's not free, but it's only like $20 or something. Does much more as well.

[Nyobie]

This forum never ceases to amaze me with how much knowledge is out there collectively. You guys are great - thanks for sharing! I have a BSEE and I'm able to typically follow the technical explanations without any problem, provided the writer follows a train of thought that's logical and I absolutely LOVE technical write-ups that are this in depth.

Keep writing! (Where's that little 'Respect!' icon when ya need it?)

[_GUI_]

Increasing the ISO will decrease the signal to noise ratio...usually.

Output SNR improves the higher the ISO set. For a given amount of light entering the sensor (i.e. if the aperture/shutter combination remains constant), raising ISO produces higher output SNR because it provoques the noise entered in the ADC stage (Npost) to contribute less to SNR:

http://www.guillermoluijk.com/article/iso/modelo.gif

So as long as you don't clip the highlights, for a given aperture/shutter shoot at the highest ISO possible:

(all shots at the same aperture/shutter):
http://www.guillermoluijk.com/article/iso/versus.jpg

Always talking about real ISOs, i.e. those which actually correspond to analog amplification values (generally up to ISO1600 in most DSLR's and ISO3200 in the newest cameras. ISO6400 on a Canon 5D2 is a fake ISO, the highest real ISO is ISO3200).


OK. But which will produce less noise, a shot at f/2.8 ISO 6400 for 1/2000 or the equivalent shot of f/2.8, ISO 100, 1/30?.

The shot at 1/30 will produce far less noise (higher SNR) since there is much more light entering the camera.

Regards.

[Scottes]

Guillermo, you seem to be contradicting yourself. You say to shoot at the highest ISO possible, then you quote my example and say to use the lower ISO.


But... your example shots... Are those both at the same aperture and shutter speed, like you used an ND filter on the ISO 1600 shot?


Maybe I just need more coffee....

[_GUI_]

Guillermo, you seem to be contradicting yourself. You say to shoot at the highest ISO possible, then you quote my example and say to use the lower ISO.

But... your example shots... Are those both at the same aperture and shutter speed, like you used an ND filter on the ISO 1600 shot?

Nope, that is why I insisted in: for a given aperture/shutter. Your question was a different case since you changed shutter (1/30 vs 1/2000). In that case is much better the longest exposure because it will allow more light impacting the sensor, no matter the ISO.

The sample shots were not done using any filter, exposure was adjusted in pp to be the same (obviously the RAW file at ISO100 was strongly underexposed because aperture/shutter was kept the same).

The conclusion is simple: if you don't have light enough, don't be afraid to push ISO as much as you can before clipping the highlights to reduce final visible noise. But if you can expose well at base ISO, it will be best to stay at base ISO because you'll have more light entering the camera and hence better SNR.

Take your coffee and read again :D

[Scottes]

OK, I'm more awake now. :D

I'm not sure your example is fair though. I would think that pushing the exposure in LR or PS is different from how the camera does it.

That is, given the following two situations:
ISO 1600, f/8, 1/30 with a 4-stop ND filter
ISO 100, f/8, 1/30 without any filter
Both shots are equivalent exposures, but I'd bet that the ISO 100 comes out better. No proof, just a theory.


I will agree that pushing an exposure in PP yields a worse result than using a higher ISO in the first place.
ISO 1600, f/8, 1/30
ISO 100, f/8, 1/30 pushed 4 stops in post-processing.
Both shots are equivalent exposures, but the ISO 1600 comes out better.

[_GUI_]

OK, I'm more awake now. :D

I'm not sure your example is fair though. I would think that pushing the exposure in LR or PS is different from how the camera does it.

That is, given the following two situations:
ISO 1600, f/8, 1/30 with a 4-stop ND filter
ISO 100, f/8, 1/30 without any filter
Both shots are equivalent exposures, but I'd bet that the ISO 100 comes out better. No proof, just a theory.

I agree talking about pushing ISO as a way to reduce noise can be quite confusing if not put into proper context. My samples actually come from a wider article which started from the constraint that we don't have light enough to properly expose at ISO100. It is in that situation when the user shouldn't be afraid of pushing ISO as much as possible. Using high ISO values will mean noise, but images will be even noisier if we stay at underexposed ISO100 shots.

In your example, noise would be much lower in the ISO100 shot since there is 4 times more light entering the camera that at ISO1600+filter.

To minimise noise:
1. Always expose the RAW as much as possible (ideally right before starting to clip the highlights, i.e. ETTR)
2. According to available light, achieve 1 at the lowest possible ISO from all the real ISO your camera provides (in general going beyond ISO1600 becomes useless on most cameras, so we can stop there even if ETTR was not reached)

Note that condition 2 is equivalent to allowing as much light entering your camera as possible.

BR

[Scottes]

To minimise noise:
1. Always expose the RAW as much as possible (ideally right before starting to clip the highlights, i.e. ETTR)
2. According to available light, achieve 1 at the lowest possible ISO from all the real ISO your camera provides (in general going beyond ISO1600 becomes useless on most cameras, so we can stop there even if ETTR was not reached)

Note that condition 2 is equivalent to allowing as much light entering your camera as possible.
I agree wholeheartedly.

[grandad35]

I have often wondered about the internal workings of the imaging electronics, and have greatly enjoyed this discussion. Reading all of this has raised a few additional questions, however.

The newer cameras have 14 bit converters, touted as a big improvement over 12 bit converters. I downloaded one of GUI’s images in Post#20, converted it to grayscale and looked at the data in an area where the colors appear to be fairly consistent. The histogram shows a standard deviation of 3.77, or approximately +/- 11 counts peak-to-peak (+/- 3 sigma for a normally distributed data set). In case anyone is wondering, the individual R/G/B standard deviations are 2.9/5.9/14.6 in the color image. Let’s be conservative, and attribute only +/- 4 of the +/- 11 counts to sensor noise, leaving an accuracy of 5 bits in the mid-tones. Allowing 4 additional bits for the anticipated improvement in an ISO 100 image (compared to ISO 1600) and 2 additional bits for the linear-to-gamma conversion in the mid-tones, that gives an 11 bit noise-free linear (RAW) signal at ISO 100 after all is said and done. GUI - Was this shot taken on a 12 bit or a 14 bit camera?

If this analysis is correct, the question is whether a 14 bit converter really makes a difference, or does it just give us more precise values for the random noise component? Have the newer cameras also improved the S/N ratio of the various analog components so that the extra 2 bits represent useful data? Is this what has allowed the maximum ISO values to be increased by 3 stops (bits) from 3200 to 25600 on some of the newer cameras?

The next questions don’t relate to noise, but concern technical details of the data conversion process. If they are too far OT, ignore them.
Is there one A/D converter per row of sensors on a Canon DSLR?
What is the approximate conversion time for a single reading ?
What type of A/D converter is used (Flash, Successive Approximation, etc.)?
The conversion process seems simple enough when there is a shutter to expose every sensor element for exactly the same time, but it must get interesting on a P&S or when using Live View” on short exposures where the total image conversion time can be on the order of the exposure time. Do they reset a column of sensors at a time, and then reset each subsequent column at the same time period as the conversion process will take to get the same exposure time across the sensor?
I seem to recall reading that the absolute lower limit on the sensor element size is the wavelength of light. Is this true?

[Scottes]

14-bit gives you more possible discrete tonal values, and that's true across any range of the image. So if you analyzed the "middle stop" of an image from a 14-bit camera, there can be more "different colors" (aka "discrete tonal values") in the same "middle stop" of an image from a 12-bit camera.

I think that the only impact this has on noise is that the noise can be more colorful. :D

[_GUI_]

Have the newer cameras also improved the S/N ratio of the various analog components so that the extra 2 bits represent useful data? Is this what has allowed the maximum ISO values to be increased by 3 stops (bits) from 3200 to 25600 on some of the newer cameras?

A very good question. The answer is not yet. According to real noise measures, only the Nikon D3X is today a sufficiently low-noise camera at base ISO to enjoy those extra 2 bits. The rest would be equally well served with 12-bit RAW files.

In another forum I asked someone to decode a 14-bit RAW file as if it were 12-bit (just need to fiddle a bit in DCRAW's source code and make a 2 bit rounding to the undemosaiced Bayer data prior to demosaicing). The result was some pixels differ in the shadows, but none of the image could be qualified as 'higher quality'.

As long as noise hides the extra tonal richness provided by an increased bitdepth, it's useless to add more bits to the ADC.

Just a sample image demonstrating more levels is useless in presence of noise higher than the levels gap: the following 2 images have very different histograms, but both display the same quality and are equally robust against postprocessing (the image on the right won't posterize before the one on the left after heavy pp):

http://www.guillermoluijk.com/article/ettr3/poster.gif

[Scottes]

Guillermo, you're being a bit harsh saying that 14-bit is useless compared to 12-bit. That's like saying that 12-bit is useless compared to 10, and 10 to 8, and so on and on until we find that there's no difference between closing your eyes versus opening them.

Every little bit helps. (Or 2 bits in this case.) The more they keep improving, the closer the come to a 32-bit format, and/or becoming more capable of capturing 20 EV of dynamic range.


I will admit that a small percentage of people can tell the difference between a 12-bit capture and a 14-bit one. So, perceptually, it's not very useful. But "useless" is a bit harsh and exaggerated.

[_GUI_]

Guillermo, you're being a bit harsh saying that 14-bit is useless compared to 12-bit. That's like saying that 12-bit is useless compared to 10, and 10 to 8, and so on and on until we find that there's no difference between closing your eyes versus opening them.

Every little bit helps. (Or 2 bits in this case.) The more they keep improving, the closer the come to a 32-bit format, and/or becoming more capable of capturing 20 EV of dynamic range.

No, it is not the same at all. Noise in present digital sensors is much bigger than the gap between two levels in a 14-bit scale, it is bigger than the gap between two levels in a 12-bit scale, but it becomes not so big compared to the gap between two levels in a 10-bit scale (BTW most compact cameras have 10-bit RAW files, because they actually couldn't benefit from more bits). That is why 12-bit are useful today compared to 10-bit. But 14-bit are a waste of resources compared to 12-bit (I insist, in present digital cameras. In the future this will change of course, and more bit will be necessary).

A given bitdepth in the ADC is useful only if noise is low enough to enjoy the extra levels. If this is not the case more bits are unnecesary.

Bitdepth is not the limiting factor of DR, noise is. You can change the ADC on any 14-bit camera by a 32-bit ADC; as long as the new ADC is as noisy as the old one, DR and quality of results won't improve at all. Imagine that you are measuring the distance between 2 cities with a device that measured 200,4327 Km. But this device has an error margin of up to +/- 1Km. Do you think the figure 200,4327 Km is more precise than the measure 200,433 Km?. It's the same with noise and digital sensors.

Emil Martinec has studied this topic in his brilliant article: Noise, Dynamic Range and Bit Depth in Digital SLRs (http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/index.html). The 14-bit question in particular is studied in Noise, Dynamic Range, and Bit Depth (http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/noise-p3.html#bitdepth):

"Both Canon and Nikon have introduced a finer level quantization of the sensor signal in digitizing and recording the raw data, passing from 12-bit tonal gradation in older models to 14-bit tonal depth in newer models. A priori, one might expect this transition to bring an improvement in image quality -- after all, doesn't 14-bit data have over four times the levels (16384) compared to 12-bit data (4096)? It would seem obvious that 14-bit tonal depth would allow for smoother tonal transitions, and perhaps less possibility of posterization. Well, those expectations are unmet, and the culprit is noise."

Regards.

[Scottes]

I will have to read those articles when my brain is not occupied (burdened) by work. Thanks.