Noise model for datasets with min/max datapoints

[michael-sutherland]

I have a noisy dateset from an SEM that has a lot of pixels at 0 and 65535 (it is a 16-bit dateset). For these pixels, the noise response is a little special. I would assume they wouldn't change value in most cases. Also, they aren't a good candidate for pixel replacement for N2V or PPN2V (especially as the pixel being picked as a neighbor pixel to replace an existing pixel). Do you have any suggestions on how to configure N2V and PPN2V to best fit this type of image? Or algorithm changes that would make sense in this case?

Here's what the noise histogram looks like in the image, with the edges still there. Obviously, the GMM fits a lot better if we trim off the extremes, but has issues correcting noise, I think because the noise modeling for these pixels doesn't really match the rest of the pixels.