P: - Add ISO and Exposure Time to the Noise Profile Tag implementation

For image processing algorithms, reading the noise profile is a very efficient way to caracterise the noise without doing a costly noise estimation. This is very usefull for inboard algorithms. If the noise profile values are not properly defined regarding the ISO and exposure, the noise profile is very innacurate and cannot be used. For multi-exposure burst processing in particular (such as HDR), this is a real drawback.

Using a little contextual noise reduction here but distilled to its essence you are asking for:

- For which ISO is the noise profile tag given

- For which exposure time is the noise profile given

Would you like me to conver this to a feature request for you?

I am adding explainations to my original post, since it seemed to lack clarity.

Cameras are affected by both read noise and shot noise. The variance (or power) of noise is commonly expressed as : 

Where I is the pixel brightness (between 0 and 1).

By considering G the ISO gain, P the photon flux, t the exposure time and Qe the quantum efficieny, the brightness can be expressed as :

The read noise is often considered to be independant of brightness, ISO and exposure. The shot noise is modelled as a Poissonian noise on both photons and electrons, amplified. By calling D the dark current, this yields :

As a result, the total noise power is often considered to be following the law:

The noise power is therefore often considered to be growing linearly with the brightness, with a slope and intercept alpha and beta, dependant on the ISO, exposure and camera. Alpha and Beta are called "noise profile" and are expected to be found in one tag of the .dng files. They are not estimated from these equations but rather from noise estimation methods.

Do note that the noise power is not a good quantity to asses the "noisyness" of an image, but rather the signal to noise ratio (SNR):

Now that all this is cleared, my remark was the following :

From the equations, alpha and beta are expected to scale with the ISO gain, and when reading the tags, it is indeed the case.

However, beta is expected to scale with the exposure time too (because of the dark current), but the value in the tags does not. When performing my own noise estimation, it is indeed scaling as expected. I suspect that the tags where tabulated for different ISOs, but a single (unknown) exposure. What should be the convention of the .dng format on that ?

I don't think it has anything to do with ISO invariance; from what I understand ISO invariance would translate to sigma read = 0, which would actually make the model more accurate.

Regardless of that, not every device is ISO invariant, and having such an ambiguity on the given noise profile tag somehow makes it unusable. If you believe that this model is inaccurate or outdated, please point out what makes you think so.

Here are a few (rock solid) references to back up the equations :

https://graphics.stanford.edu/courses/cs448a-10/sensors-noise-14jan10-opt.pdf

https://people.csail.mit.edu/hasinoff/pubs/hasinoff-hdrnoise-2010.pdf

Digital cameras produce both shot noise and read noise, and a lot of cameras today approach ISO-invariant. So I'm not sure how much of the above is accurate anymore.