Quote:
Originally Posted by Shiraz
But I think that when you do the subtraction processes, you include noise from 2 or more pixels, rather than just one (as for each RGB pixel) - so you may end up slightly worse off in the end (or at least no better off). eg if you had a large noise excursion in your L channel, you would end up with a large noise excursion added to all 3 RGB channels as well.
The other issue may be dynamic range. I seem to recall reading that the Kodak scheme of using clear L pixels in a modified Bayer matrix ran into trouble with the L pixels filling up long before the colour ones did, so the exposure had to be truncated before the colour data had got very far above the noise - nice L but poorer quality RGB. Same thing could possibly apply here.
and then again I could be completely wrong  |
Interesting points! I've done some experimentation, which is probably not mathematically sound :p
I generated an L frame in PixInsight, from which I extracted RGB of varying percentages (30, 50, 20 respectively).
Onto the L frame I added some poisson noise with an amplitude of 0.6. Onto each of the RGB I added 0.2
From the RGB I generated, by simple summation, P and Y frames. From the P and Y images I then reverse-engineered the RGB frames as I've proposed above, with results as attached.
At this point I basically have no idea how to evaluate the results. I assume that a higher variance means there's more noise. If it's as simple as that, then you'll see that the LPY version has more noise in RG but less in B, with more overall (0.01e-03).
Even if I didn't stuff up something obvious this isn't reflective of reality because in reality the LPY would benefit from reduced read/bias/dark noise whereas this experiment assumes that all noise is doubled when combining the RB and RG images.