decay wrote: ↑Fri Oct 21, 2022 6:06 am
I’ve started thinking about what this means in case of noisy data sets.
If I understand correctly this means for every single pixel
- we have two signals (original or accented) with different levels
Correct.
- both signal levels are superposed with different, random noise levels (shot noise or whatever)
Correct.
- therefore the outcome (which signal “wins”) relies on the (for each pixel) concrete, random, statistical noise levels both signal levels are superposed with.
Correct.
This means if there is for example an area in original data with signal level(s) not too much above the signal level of the accented data and both signals are noisy that there are a number of pixels in this area where the accented data “wins” – due to random noise levels. So the accented data shines or pokes through – as you wrote.
Do I understand this correctly? (I hope what I wrote was halfway understandable. Not so easy to describe.)
Correct. Of course, the signal is first transformed (stretched, thresholded)
and modulated in various other ways by the non-accented signal.
Having all this said, here comes my question: Would it make sense to do
first noise reduction on both data sets (separately) and than
afterwards NB accent processing as the ST module does? Having lower noise levels on both data sets should reduce this ‘poking through’ effect?
At the very start of NBAccent's research and development, that's what I thought as well.
It turns out though, that Tracking does a
much better job on the combined signal during final noise reduction, rather than on just the NB signal before compositing in the NBAccent module. It makes sense when you think about it; the signal flow is setup such that Tracking regards the NB signal as "not there" in the final image, meaning that any shot noise (e.g. non-correlated noise) that is introduced by the accents is completely obliterated in the final noise reduction stage, if it thinks the non-accented signal/detail suffers. The accents are
truly in the service of the original detail from a signal processing point of view.
This is also the reason why using the NBAccent module should yield much better
final results than any traditional compositing technique, and why you should be able to use
very marginal data to enhance your visual spectrum image with accents.
The latter is very often the case when dealing with Ha accents of galaxies (as I mentioned in that post); there is usually very little signal except in a few "knots".
For a good example of such marginal data see the
Elf's Cloudy Nights data. One of my most challenging test datasets was his Pinwheel galaxy. Visual spectrum (intentionally left as luminance only) looks like this;
-
- Selection_756.jpg (208.74 KiB) Viewed 14942 times
But the Ha data looks like this;
-
- Selection_755.jpg (534.83 KiB) Viewed 14942 times
The background is very uneven and thresholding is needed. All said and done though, after compositing noise is very much under control;
-
- Selection_760.jpg (204.13 KiB) Viewed 14942 times
and final noise reduction removes any remnant noise very effectively as described above;
-
- Selection_758.jpg (104.37 KiB) Viewed 14942 times
-
- Selection_757.jpg (48.85 KiB) Viewed 14942 times
You can clearly see that if there is no detail in the original signal in a location, then noise reduction all but obliterates the noise blotches etc. from the NB accent signal.
Hope that makes sense and helps!
