Are there known techniques to estimate the SNR of an image in small neighborhoods (say like 11x11) without any knowledge of the clean image, with arbitrary content ? Denoising is not required.
Update:
On second thoughts, I removed the qualifier locally, as it seems superfluous.
Without any knowledge of the clean image, with completely arbitrary image content?
No. Because what if the image is a perfect portrait of noise? Without prior knowledge that an image is somehow "not noise", you can't distinguish between an image of noise (which looks just like noise to the image consumer), and a noisy image (which looks just like noise to the image consumer).
You need some prior knowledge of the image characteristics, if not of the image itself.
If, for example, I give you an image and I promise you that because it was taken on a foggy day it has a certain known degree of Gaussian blur, then you can take a noisy image of that scene and -- to a certain extent -- infer how much of the image must be noise.
As a mathematically more difficult, but conceptually easier example, if I give you an image and tell you that it was painted by some 1950's modern artist who only painted large monochromatic blocks, then you could identify the boundaries of the blocks, segment the image into "this is all a certain shade of red", "this is all a certain shade of ochre", etc., then you could, again, determine the levels of noise within those segments and, thus, infer the level of noise in the overall image.
But -- to pound home the point -- if I gave you an image that some modern artist from 1990 composed by generating random noise and rendering it into a file and said "here's an image of totally random noise that's been corrupted by more totally random noise" then there would be no mathematical process that would separate out the intended image (which is just a sample of noise) from the resulting image (which, again, is just a sample of noise).
Do you have access to raw sensor readout, or only a processed image?
Looking at raw camera sensels, I would think that assuming that the image noise is e.g. dominated by signal dependent but otherwise spatially invariant shot noise, that you could do some kind of modelling. Neighbour pixels often tend to be quite similar (particular in out of focus or smooth areas). If they are dissimilar, one component should be shot noise. Shot noise should be a global function of pixel magnitude, so the consistently «signal level dependent» component of neighbourhood variability could perhaps be attributed to shot noise.
