I want to address your desire to compute the local noise in the image: Perhaps you can succeed in the technique you are trying but i think there is a more straightforward analysis that will be helpful. First, the noise in the sensels is independent, from sensel to sensel, if you can work on the raw values. The key is not to convolve adjacent sensel values. Simply record about 20 identical images and compute the noise at each voxel as the sample standard deviation using Matlab. Hope this helps.
I'm afraid that I've caused confusion by using the term "rms noise" in the original post, and I apologize for that. What I meant by that term is the same algorithm used to calculate the standard deviation. Basically, it's the rms value of the deviations from the mean. Engineers often refer to that calculation as rms noise, even when the thing causing the deviations isn't, strictly speaking, stochastic. That's the way I was using the term, and it was sloppy, and it gave you and others the wrong impression. Another way to think of the calculation is to use a term more often used in one-dimensional signal processing: the rms value of the ac component.
In the images that I am processing, most of the variation from the mean value is signal: that is, it is caused by the target, as interpreted by the imaging system. I am well away from the read noise, so banding is not an issue. However, there is one source of real noise: photon/shot noise. There is another source of what can be thought of as noise: pixel response non-uniformity. The first can be averaged out as you suggest. The second cannot. I am operating in a portion of the dynamic range of the camera where PRNU is at least as large as photon noise.
You got me thinking about this, so I did some testing. I averaged the ten exposure-corrected images in each set, and compared the standard deviation of the averaged image to the average standard deviation of the ten individual images. The errors ranged from 0.05% to 0.14% in the eight-set aperture series I used for this experiment.
I could come up with a way to reduce the effect of the photon noise by computing the gain factor from sensor electrons to ADC counts (camera ISO / Unity Gain ISO), computing the photon noise standard deviation as the square root of the average sensor electrons count, and subtracting that in quadrature from the standard deviation I've already computed. I'm resisting that, because it seems unnecessary, and because it will make the whole procedure more complicated to explain.
I could also compute a correcting image for each camera by making a bunch of images of a flat field, applying lighting correction, and averaging to get a pixel response map which I could use to correct for PRNU. I am resisting that for the same reasons as above.
Thanks for your interest, and, again, I apologize for the confusion.
Jim