Dithering to prevent colour noise

[vlaiv]

2 hours ago, IanL said:

Might be worth having a look at this if you're using PI, or even if you're not, to understand the general principles of weighting.

https://www.lightvortexastronomy.com/tutorial-pre-processing-calibrating-and-stacking-images-in-pixinsight.html#Section5

Bottom line is that if you combine two stacked masters containing 20 subs each, simplistically you'd weight them 50:50, but what if 10 of the subs in the first master are "poor" and only 2 of the subs in the second master are "poor"? Would you now think 50:50 is the correct weighting? The only practical way to ensure correct weighting is to stack everything in one go.

Yes, only problem with this approach is that you assign single weight per sub and that is not optimum solution.

Level of noise consists of multiple components - one of them being shot noise whose level depends on signal. Imagine following scenario - there is one frame that is lower in signal by 10% due to poor transparency, and other frame that has a bit more light pollution.

Once you equalize frames, you end up with same amount of noise in the background where is no signal (only LP noise and read / thermal noise), but because signal was lower by 10% and you had to scale it to equalize it - so does you scale shot noise (and LP noise and read noise - but since LP was higher in first image these turn out equal where there is no signal).

Now you have situation where background weights should be 1/2 : 1/2 because level of background noise is equal, but in signal area weights are different - might be something like 1/2.1 : 1/1.905 (or whatever proper ratio is to get total of 1).

In general case you want to assign weights based on signal level in particular area, and also in patches of image (because LP can be gradient and contribute different levels of noise to different parts of image) - and this is not really something that is easy to do.

I've developed signal based weights estimation and it works really good. You can assign "bins" like 16 or 24 different signal level bins, and all the pixels mapped to each bin get their weight and participate in their own "stack zone".