Yes, we take pictures of real scenes, not flat fields.
With this understanding we can assume that experiments done using flat fields are not fully describing natural image situations.
Since the signal:noise varies with photon flux, the flux has to be held constant for a proper experiment and we have to sample enough photons to get a valid sample size so as to yield a statistically valid standard deviation.
For a flat field that makes sense. But the issue in the reality is that we have only a single photo and how do we define a measure of SNR, especially one coming from the shot noise? Classroom experiments and displaying power spectrum of flat fields, as many have done, does not answer that question, except reproducing well-known facts regarding white noise known in the signal processing for decades. At this stage we must have an awareness that if we are to tackle the reality of natural images, as in photography, then some of the basic models need to be developed further. It is possible that the situation may become too complicated, and some assumptions have to be put in. But, at the first stage there should be a realization that flat fields are not we are looking for.
A real scene could be broken down into regions and the standard deviation obtained by the process of integration.
Are you saying that you would quote me 5, 10, 20, ..., numbers, complete with area coordinates from which the respective SNRs in the seemingly uniform patches were derived, regarding a simple question if I ask what is the measure noise in this picture that I took with my favorite camera?
Shot noise is the predominant source of noise for all but the darkest regions of the image, where read predominates (ignoring thermal noise and PRNU), and shot noise does follow a Poisson distribution.
The operative word I used in my original message is the "usual" Poisson process, as done in flat fields. A natural image may still be described by a more complex Poisson process, with some very interesting properties.
DXO does give a full SNR plot and also a value for 18%, which is often taken to represent average scene reflectance. I really don't get your point.
Here is the point: What exactly is the meaning of standard deviation of shot noise being the square root of mean signal in digital imaging? As my example in the following links shows it is only applicable to group of images:
http://www.luminous-landscape.com/forum/index.php?topic=51782.msg427055#msg427055Not applicable to a single image, in general, which is usual photography - does one acquire 100 images of the same static scene just to get a better handle on shot noise? What is weakly applicable in the case of a single image is that the sqrt of a pixel electron/photon count gives an approximate handle on the standard deviation of noise, if the count is high enough. But this number is not the actual noise value on that pixel in a single picture, as sqrt is an average measure of noise if a large number of pictures had been acquired. Some interpretation of the validity of the sqrt of pixel count as being shot noise, which is almost always done without even the blink of an eye, is needed. Under what circumstances is that valid? For a single image that will basically boil down to transitioning from a temporal statistic to area statistic (i.e. spatial) in the neighborhood of a pixel. For example, if the signal variation in a pixel neighborhood is smooth enough, then what is the equivalence of pixel sqrt count to the area statistics. This issue is not as trivial as it is usually treated.
Sincerely,
Joofa