TriX Loc: Raleigh, NC

Question: which of the methods of amplification is used by the major camera manufacturers - specifically Canon, Nikon and...). Thanks again for the information

selmslie Loc: Fernandina Beach, FL, USA

Probably both.

There is likely some adjustment to the reading of the physical analog signal recorded on the sensor in order to ensure linearity during the pre-ADC conversion.

There is then a conversion of this adjusted analog signal that involves a simple multiplication or gain (post-ADC amp) as it is converted to the digital values to be stored in the raw file. This is essentially the same simple adjustment used for the EC adjustment in the editor.

selmslie Loc: Fernandina Beach, FL, USA

To be more specific, things are not as simple as we would like to think they are. Our understanding of what Nikon, Sony, etc., are doing may be too superficial.

In theory the sensor is simply recording a linear sequence of tonal values in proportion to the photons received that double with each additional stop of exposure until the sensor's buckets are full to capacity. To make these into a visible image the linear sequence needs to ultimately get converted to a logarithmic scale. But when does this take place?

A sensor may not actually record a voltage in precisely direct proportion to the number of photons that are captured. The process of reading that voltage may not produce a precisely proportional numeric value - it may take a little fudging to get what is needed.

Take a look at my earlier post where I presented a case in which the scene's dynamic range was probably 6 or more stops. On top of that, I added both +3 and -3 stops so that the total range recorded in the raw file was 12 or more stops. After adjustment in PP, it is very difficult to tell the two images apart. How is that possible? How do you successfully record a 12 or 13 step range of values using only 14 bits? In the darkest tones you would have to be running out of bits and it should be easily seen in the -3 EV sample.

I think the explanation lies in the way that the raw file is generated. It makes no sense to use half of the tonal values on the brightest one stop. If they did that, they would certainly run out of bits 12 steps down. What is more, it would be impossible for the sensor to posses a dynamic range of more than 14 stops if all you had to record it with is 14 bits.

The only way I can explain both paradoxes is that the analog to digital conversion actually creates a raw file that is not as simple as we imagine. At least some of the linear-logarithmic conversion is already taking place. This is likely when the raw file is not loss-less and it may even be true when itis absolutely lossless.

Nevertheless, we don't need to know exactly how all of this is accomplished unless we are working for one of the manufacturers. I could be wrong, but the examples I posted say otherwise.

TriX Loc: Raleigh, NC

[quote=selmslie]Probably both.

There is likely some adjustment to the reading of the physical analog signal recorded on the sensor in order to ensure linearity during the pre-ADC...

Thanks for the insights. Actually the reason I'm wondering how various manufacturers implement variable ISO control is I'm still trying to get my brain around the idea of dramatically under exposing at "native" ISO and fixing in post as opposed to changing the ISO to expose "correctly" as long as you don't blow out the highlights in either case. It seems to me that dynamic range in a data acquisition system, whether audio or imagery, is ultimately controlled by the dynamic range of the A/D (bits or resolution and linierity as aperture uncertainty isn't really applicable in single-shot acquisitions) and the noise floor of the signal. Severe under exposure would seem to "throw away" the MSBs of the A/D, diminishing dynamic range. Then multiplying the digitized value either post A/D (iF controlling ISO in that manner) or in PP also increases the noise by the amount of the multiplication. If on the other hand, you apply enough amplification to the analog signal (to control ISO) prior to the A/D to come close to saturating the MSB, you utilize the entire dynamic range of the A/D, but you're amplifying the noise also, which may restrict dynamic range as the noise may take up 1 or more of the LSBs. The reason I'm concerned about how each manufacturers control ISO, is that for example, Canon and Nikons typically behave differently in respect to dynamic range vs ISO, and I'm wondering if it's completely due to the sensor noise or whether it's due to how they control ISO. If for example, you pick something like a 5D3, the noise and dynamic range is worse then an equivalent Nikon D600 at "base" ISO of 100, whereas the Canon actually has superior dynamic range at higher ISOs above 6400 (see graph below). That would seem to suggest that using the Nikon to it's best means shooting at lower ISO, (even though underexposing) and fixing in post (as has been suggested), while with the Canon would perform better by shooting at the "correct" metered ISO with less correction in post. What are your thoughts on this?

Apaflo Loc: Anchorage, Alaska

Two things are certain: it does not happen while the data is in an analog form, and the other certainty is exactly the way it is done in the step that converts from raw sensor data to image data. Every "raw converter" does it. They all do that in virtually the same way.

There is a sensor specific color correction curve applied at the same time, but the gamma correction is in no way as mysterious as you are making it out to be.


Cite an authoritative source that confirms what you are claiming.


A 14 step range is exactly what a 14-bit depth linear ADC can record. The maximum possible SNR will be

SNR = (6.02 N) + 1.76 dB

Where N is the bit depth.

Note that there are 6 dB per fstop.


Where is the paradox? A 14 bit linear ADC does in fact only record 14 stops, it does in fact "run out of bits" 12 steps down (due to quantization distortion). There is no gamma correction in the analog portion of the signal path.

It isn't at all obvious what you mean by "lossless" in that statement, as there is no discussion of compression. Any compression, just like any gamma correction, necessarily is done to the digital data and never to data in the analog state before the ADC.

ISO amplification is an analog operation. Any adjustment to the image brightness that is done to digital data is not really a change to ISO, just to brightness. Which is to say that regardless of what the slider may be labeled in this editor or that editor, ISO cannot be changed after the ADC (i.e., in an editor).


Your examples don't have much significance. In fact we know in general how the manufacturers are doing things. We don't know all the specific details for any newer model camera, as it takes time to work out how it is done. Within a relatively short time we do know exactly. The biggest problem for casually interested photographers is random "examples" posted with invalid commentary about what supposedly can be observed...

blackest Loc: Ireland

Have you had a look at in camera development of raw files?

There are a number of parameters available :-
1)custom image: this gives the usual normal vivid b&W modes plus saturation, hue, high/low key adjust,contrast, sharpening
2)White balance:
3)ISO: +/- 2 stops
4)ISO NR: High-iso NR
5)Shadow Correction: off 1,2,3
6-7)two more settings grayed out that one i think is sensor angle
and finally
8)color space: adobe RGB or SRGB.

This is on a Pentax K5 your camera may have different options. So what is the Iso adjustment in camera?
It appears at least some of the signal conditioning is post ADC at least with the Pentax K5.

Apaflo Loc: Anchorage, Alaska

Not true. That would be an exceedingly complex operation, and the primary effect would be more noise and a vastly slower camera.


"ISO Invariance" is a mythical solution in search of a problem.

Read what you write in the next paragraph! It has become clear to me over several of your posts that you have significant experience with the recording of analog data using digital systems, and the illogic to various misunderstandings is pretty obvious to you...


Looks like you have a good grip on basics! It might help a little to read the basis for all of this. Claude E. Shannon wrote "A Mathematical Theory of Communications", and started the Information Age, back in the late 1940's. The question then was what is the difference between data processing with analog data and digital data. Which is more efficient and which is more effective. Before WWII virtually all of these digital methods had been worked out in theory, but in practice it took a room full of electron tubes to do practical tests. Nobody was really sure where to invest money! Shannon answered that, and everything since then has centered on digital technology. It took until the late 1960's for the needed technologies to catch up (solid state electronics primarily), and digital has dominated everything since.


First, the DXO measurementa are fine, but the graphs are very misleading. Rather than use the cited graph, take a look at what Bill Claff (user bclaff on UHH) produces. This graph is specifically for the same two cameras,

http://apaflo.com/misc/uhh/chart200.jpg

Note the smooth graph for the D600 with an almost straight line slope from ISO 100 to above ISO 12,800. That is "ISO Invariance". Note the bumps at ISO's lower than 3200 in the 5DIII graph. That is "ISO variance". There certainly is a lesson to learn there, and you are right about where the source of that characteristic originates. The lesson is not that one can under expose by 3 to 6 fstops with a D600 and expect to recover in an editor the same image as a proper exposure. That would be like hiding behind the nearest tree in a forest and claiming to your companions by cell phone that you actually aren't there. The tree in this case is that a print doesn't have enough dynamic range to show the effect (as long as the image doesn't need any gamma expansion/compression). If the brightness corrections is the only needed edit, that tree will hide you. If the image would benefit from pulling up the shadows, or any other contrast manipulation, forget it!

The cause of the bumps vs straight line in the ISO graphs has to do with exactly where and how the camera adjusts ISO. Virtually all modern cameras use analog gain between the sensor and the ADC, and perhaps also adjustment of the ADC's DC comparator voltage, to change ISO. Most cameras also have some form of pseudo ISO (lower quality that may not be acceptable for every use) adjustment using digital manipulation. But for this discussion only the analog gain is of interest, because the digital gain, whether done in the camera or in an external editor, is inferior and should be avoided to the degree possible.

The lumps in the 5DIII chart (similar by the way to a Nikon D5 chart), are because designing an analog amplifier with sufficient gain and dynamic range with a low enough noise figure is clearly not yet possible. The 5DIII and D5 are designed with a good noise figure, to allow better results at high gain. The D600 (and D750 and D8xx models) don't have that great noise figure, but do have the dynamic range to deal with maximum output from these sensors. The bumps (in these and in other sensor curves, where they may be just one or two bumps and may be at higher ISO values) happen when the camera switches between different forms of ISO gain. The bumps at low ISO values are usually from switching between a digital controller on the amplifier to a digital controller for a loss device (a pad). The amplifier steps between 1x, 2x, and 4x gain. The pad provides 1/3 fstop steps between the amplifier steps.

The steps you might see in cameras like the D810, which are one or two bumps at higher ISO values are where the gain adjustments change from using amplifier gain to changing the DC voltage on the ADC.

Apaflo Loc: Anchorage, Alaska

Yes, but that is off topic in this thread. The topic is how ISO is varied, and specifically what is done with raw sensor data before it is saved in the RAW file.

Everything you list is post raw sensor data. All of those items are adjustments to the RGB image data produced by demosiacing the raw sensor data. Very interesting stuff! But this is the wrong thread...

selmslie Loc: Fernandina Beach, FL, USA

I have wondered about that. Canon's approach is certainly different from Nikon's so we can't really make blanket statements about all manufacturers.

For a D810, a dynamic range of 14.4 at ISO 100 drops to 6.4 at ISO 51200. That's an almost uniform drop of 8 stops of DR over a 9 stop drop in exposure.

At the same time, S/N drops from about 45 dB to 17 dB or 28 dB over the same 9 stop drop in exposure and this drop is similarly uniform. This is pretty close to 3 dB per stop the way DxOMark defines dB.

Canon is subject to the same physical constraints but the figures for the EOS 7D Mark II suggest that its native ISO is actually greater than 100 since the DR for 200 is the same as is it at 100.

The only practical solution is test your own camera for its limitations and work within them.

I have done this and find that if I stay near base ISO I can be off as much as two stops in either direction from the "correct" exposure and recover in post processing. When in doubt, I bracket.

There is no authoritative source. What Nikon and Canon are doing are trade secrets. Everybody else is speculating - trying to reverse-engineer the process. We can guess but we will never know the truth and you will never be able to Google it.

Theorizing and speculating about how all of this works is a fruitless mental exercise. It might satisfy your mental self image but it does not produce any useful results.

The manufacturers are sharing their approach with Adobe and other software developers on a need-to-know basis and under strict confidentiality agreements. The software developers are not going to turn their code into shareware either.

So on both counts, manufacturers and software developers, you are simply wasting your time (and ours).

selmslie Loc: Fernandina Beach, FL, USA

The same characterization can also be applied to ETTR. "ISO Invariance" is a misnomer - a misleading label.

However, most of this discussion has been about the possibility of recovering image information from an "underexposed" image. There have been plenty of samples posted to show that you can underexpose an image and still recover shadow information that is relatively noiseless if you stay close to base ISO.

ETTR is a solution to a problem that no longer exists. It arose in the time of narrow dynamic range cameras where careful metering and evasive action was needed to deal with noise.

The difference between the proponents of ETTR and the expose to the left crowd is that the latter are willing to demonstrate their claims with real images. Proponents of ETTR are incapable of doing the same and appear unwilling to demonstrate any useful real-life examples. All we see are overexposed images brought back to "normal" in PP. There is no magic in that so long as the scene's dynamic range is narrow and you don't blow the highlights.

What both approaches have in common is that the chosen scenes have much less dynamic range than the camera is capable of capturing at the chosen ISO. If you reverse the situation, only HDR will solve the mis-match.

Apaflo Loc: Anchorage, Alaska

Asked for an authoritative source, this is the response:

In fact one can Google it. Nikon and Canon may have trade secrets, but that is not one of them. The physics of CCD and CMOS sensors are well known, and while camera manufacturers work at the bleeding edge of how to manufacture useful sensors, the actual bleeding edge for what can be done is not nearly as proprietary.

Also consider that neither Nikon nor Canon ever put anything into a camera that they have not patented. They do not attempt to keep it secret so much as simply monetize their knowledge. Nikon works closely with Sony, as an example, and commonly has agreements that allow Sony to manufacture a sensor for Nikon, and one year later lets them use the technology for their own sensors.

The fact is, there is no authoritative source for what was stated because it is totally off the wall nonsense. That is a well known fact and it can be Googled. Try searching on the words "raw sensor linear photon" to get about 1.35 million hits. None will say the sensor is anything other than linear, and that any non-linearity is purely a fault rather than a function.

Then don't continue to do it. You just obfuscate the issues.

Interestingly enough, with virtually every new camera that comes out Adobe is not the first one to have a working RAW converter. They may have NDA's and access to proprietary information, but they base the code they use on Dave Coffin's freely available DCRAW. He does not have an NDA with anyone. Adobe actually isn't the only one who does that either. Look up libraw.org to see how far it goes. RawTherapee is an example program.

Yeah, sure.

Tell us again how dB for voltage is different than dB for power... We all like laughs, but...

Apaflo Loc: Anchorage, Alaska

Actually we can. Note the similarity between the curves for a Canon 5DIII and the Nikon D5. They are doing the same thing. Nikon only needs to do that on the models intended for best and fastest operation at high ISO values. Canon can't avoid it even with cameras that are intended to be slower and not make best use of the highest ISO values.


Yes, but that doesn't explain anything about what either Canon or Nikon is or is not doing. Just citing the measured results does not answer the questions that TriX asked.

DxOMark does not define dB. They are defining their notion of Dynamic Range and they are also deciding where an how they measure SNR. They do explain their Dynamic Range, but they don't explain their idea of SNR well, and it clearly confuses a lot of people. The problem is they are comparing values that are gamma corrected and not explaining that it is not linear encoded raw sensor data, but RGB data after RAW conversion.

With linear encoded raw sensor data there is a 6.02 dB step between each fstop. If the gamma 2.2 correct data is used there is a 6.02/2.2 dB step between each fstop.

Regardless of how that is done and what it actually means, it does not answer the question that has been asked.

selmslie Loc: Fernandina Beach, FL, USA

What does that show? If I Google "floyd davidson takes great tropical pictures" I get 9.9 million hits.

That's exactly what I said. In practice, there is non-linearity and it has to be addressed either before or after the analog to digital conversion.

All of these issues are simple only if you are simple-minded. The fact remains that Canon and Nikon handle things differently. You cannot explain the difference. To you it is a mystery.

There are lots of reasons for non-linearity including operating temperature and the nature of light itself which, having different wavelengths and refraction characteristics, get recorded by the sensor in less than a precise theoretical proportion.

The individual red, green and blue filters get as close as they need to but they do not precisely isolate the wavelengths either. There is also the Foveon sensor in which they are stacked. Maybe you would like to ponder that mystery as well.

Maybe you would like to ask them why. Is is because they need to make accommodations for the different ADC formulas needed? Or are they simply trying to force their customers to upgrade to their latest software?

As an engineer and a mathematician (you are neither) I know that there are things happening inside the hardware and software about which neither of us can speak with ultimate authority. Since I am better equipped than you are to understand the issues and I am willing to admit that there are things I do not know, how do you think this makes you look? A bit sophomoric?

I neither need to know nor want to dwell on the details. I also don't know how Michelin compounded the tires for my car. I just keep them properly inflated and replace them when they wear out.

Tell us why you still don't know the difference between amplitude ration and power ratio. Is it because you cannot read or are you not intelligent enough to understand it?

Apaflo Loc: Anchorage, Alaska

There are not multiple definitions of dB, just one. The relation between amplitude and power is very simple: P = IE. And that is the "difference" shown in the chart at that URL.

It is not a case of two different definitions. They are equivalent; exactly the same thing!

So explain it to us again! We need more humor in this discussion, eh? Or maybe not...

selmslie Loc: Fernandina Beach, FL, USA

I guess that't the only definition you understand. Too bad!

The table is clear. A power ratio of 10 is represented by 10 dB (the classic definition) but the amplitude ratio of 10 is 20 dB. That's two definitions.

Even if you can't understand the difference it will be clear to anyone else who looks at that table and reads the article.

Maybe the rest of us are smarter than you are.