Troy,
I think you are pretty much on the money.
I would just add that the IEEE 32 bit format has a 23+1 bit mantissa, so for data with limited dynamic range, (eg a stack of unsigned 16-bit integer images) it is like an int-23 for quantisation noise, so potentially not as good as int-32, but in reality they may be pretty similar.
Definitely, a 16-bit result of stacking 16-bit subs has way more quantisation noise than either an int-32 result or a float-32 result, so I would avoid converting to int-16 until after the curves have been adjusted.
And be sure to save in a 32-bit format in case you want to come back and tweak the curves later!
-Ivan
|