Now if you really want to see how much binning reduces noise, yep it does work, here's two examples.
The first is with superpixels made by averaging each Bayer pattern, it includes two green, the red and blue.
The next is the average of 4 superpixels, with 8xG, 4xR and 4xB.
This follows on nicely from the simplified workflow I made above, because the WB and gamma are already done, it's in a proper greyscale image ready to average.
Made with
dcraw64.exe -d -w -T canon_powershot_sx50_hs_08.cr2
magick canon_powershot_sx50_hs_08.tiff -scale 50%x50% grey2.jpg
magick canon_powershot_sx50_hs_08.tiff -scale 25%x25% -resize 200%x200% grey4.jpg
The -scale command is nicely supplied by IM already:
https://www.imagemagick.org/script/command-line-options.php#scaleIf when shrinking (minifying) images the original image is some integer multiple of the new image size, the number of pixels averaged together to produce the new pixel color is the same across the whole image. This is a special case known as 'binning' and is often used as a method of reducing noise in image such as those generated by digital cameras, especially in low light conditions.
I can also add that the standard deviation of the noise (assuming it has a normal distribution, which the read noise does as that's a thermal noise (but not the light itself which has Poisson distribution), will be reduced by
sqr(8) = 2.8x
, on the green channel, or approximately the luminence noise.
So for this camera where I have measured standard deviation as 54.1 for ISO6400 (after black level subtraction), it would be reduced to 19.3, which is about the level of ISO3200.
Put another way, binning of 4 superpixels lowers noise by one stop.
I need to think about this further, because it's kinda complicated.