I believe it would be advantageous to do demosaicing/debayering after green channel registration and during the stacking process. Is there a way that Siril currently does this? If not, is it on the future feature list?
Some background - all demosaicing options (except just making a super-pixel) attempts to fill in color information through context of the pixels around it. In contrast, algorithms like drizzle attempt to use multiple sub-pixel registered frames to fill in detail data on a denser pixel grid. In the drizzle case, it actually relies on the images being slightly shifted from each-other in order to work properly.
What I’m suggesting is that these concepts should be used in tandem. Instead of guessing at colors first through demosaicing, use the drizzle results in the green frame to properly fill the red/green/blue channels with actual registered data during the stack.
As a close example, my Pentax K-1 has a “pixel shift” feature. It can take 4 “sub-images”, including a normal image, but then moving the sensor (on an electromagnetic platter) 1-pixel-width right, down, and down-right to capture 3 color-pixel shifted images. Then when it computes the final composite image, it fills in the color information on the final image with the values from each subimage. We can/should be doing the same thing for our astro stacking images.
I would do this myself by just using the pixel-shift output of my camera as the input to Siril, but unfortunately my sidereal tracking isn’t perfect, so the output isn’t perfectly aligned.
Hello.
What you are looking for is the Bayer Drizzle.
Yes we have a TODO list entry for that.
However, my last tests showed me that the improvement of such a feature was not that big. Indeed, our last debayer algorithm (especially the RCD one) are very good and even if Bayer Drizzle brings resolution, noise level is increasing.
This is why we have postponed this feature.
Thank you, Lock, for the note! Just curious, but would noise increase if the resolution is kept the same? I would understand that if we tried 2x drizzle, but I think this technique is still valid at the base resolution.
I’ll keep an eye out for RCD. I take it that’ll be 1.0 feature? Any hints on when I might see that drop?
To illustrate what I’m saying, here a small comparison between Siril and APP using Bayer Drizzle.
The only case where resolution is much better than Siril is for the Point Kernel situation.
But as you can see, the noise is (too) important.
But in some cases, Siril is even better than Bayer Drizzle.
I agree that the point kernel is the only one with a noticable resolution increase. As for noise, maybe the color halos around the bright stars are more pronounced?
I discovered last night that there’s another drawback to Bayer drizzle. At least for DSS (and I think APP), the stacking mode is limited to averaging. I prefer to use kappa-sigma to get rid of satellite trails.
Anyway, I’m looking forward to trying RCD, and especially 32bit float native support in 1.0! Thanks for the great software!
