Sampling

[vlaiv]

2 minutes ago, ONIKKINEN said:

Right, didn't read the whole discussion properly it seems. Just tested the command with already split (so mono) data and the command still runs. The files are still named CFA_0,1,2,3 but now of course they dont matter since all of them are mono anyway so one could throw them all together.

On the bin x3 issue i am experimenting on what the best way to reach that is, or one could interpret it as what the best way is to reach bin x 1.5 with OSC subs split to mono to half resolution which locks away bin x3 from the raw data but this also applies on wanting to bin 1.5x from a mono camera. I am experimenting with just resampling down to 67% (or thereabouts), debayering the data instead of splitting an then binning them and lastly simplified x2 drizzle integrating (super resolution stacking, not real drizzle in siril) the mono subs to 200% size and then binning that x3. Not sure what to make of the data i have gathered yet so wont comment anything concrete, but it looks like either the split-drizzle2x-bin x3 route or the debayer + bin x3 is the best way from the looks of it. Resampling with a smart algorithm will have good sharpness, but SNR wont increase all that much, using a simpler algorithm goes the other way. Maybe there is a middle ground algorithm that does a good job, but not sure what that would be.

Sounds like a lot of trouble to drizzle integrate thousands of subs but Siril actually runs through a mountain of subs that way in not that long. Registration and drizzle stacking more than a thousand subs is a 20-40 minute process on a decent PC. Compare that to APP (other stacking softare i use) which takes maybe 5 hours to do all that WITHOUT the drizzle.

I did some experiments with non integer binning and I have come to the following conclusion:

Just bin to nearest whole number and then up sample subs before integration using some quality resampling method.

If you want to bin x1.5 - bin x2 and then up sample by factor of x1.3333.

Resolution won't take large hit and bin x2 will improve SNR by factor of x2 - so you'll get wanted resolution and still get good SNR improvement.

(alternative is to bin and then drizzle integrate - but I think result will be more or less the same except up sampling is much faster - it can even be part of registration process as alignment of images uses resampling anyway).

 

[ONIKKINEN]

5 minutes ago, vlaiv said:

I did some experiments with non integer binning and I have come to the following conclusion:

Just bin to nearest whole number and then up sample subs before integration using some quality resampling method.

If you want to bin x1.5 - bin x2 and then up sample by factor of x1.3333.

Resolution won't take large hit and bin x2 will improve SNR by factor of x2 - so you'll get wanted resolution and still get good SNR improvement.

(alternative is to bin and then drizzle integrate - but I think result will be more or less the same except up sampling is much faster - it can even be part of registration process as alignment of images uses resampling anyway).

 

Didn't think of this one, gotta try it the next time. I am a little bit worried about the jump in resolution however, it would go from 1.5 to 3 and then back to 2.25'' so could lose some actual detail, but maybe not so worth a try.

[gorann]

Interesting, but I feel tempted to wait until someone makes a PI process that will make this for me. Will it work for both mono and OSC?

Edited November 10, 2022 by gorann

[vlaiv]

1 minute ago, gorann said:

Will it work for both mono and OSC?

For mono it works normally, but for OSC - it will work only if you first do split debayer. Regular debayering and then split binning won't give you result that you are hoping for as there is pixel to pixel correlation in interpolated values (no longer truly random - true measurements).

If you treat OSC as already having half the pixel count / half the sampling rate - then yes it will work normally after you extract color data as mono subs.