New idea - using bad sub noise

[NickK]

I've just had a bit of a brain storm. As most people throw away bad subs rather than use them. Up for a bit of a discussion...

Why not use bad sub noise to help statistically predict noise rather than throw them away, possibly allowing faint signal also to be identified in the good subs.

(1) Sub = Signal + noise

Important bit - mask out the bad signal in the sub (this can be done roughly as):

(2) SubNoise = Sub - signal

Using both good and bad subs (this is a simplistic form - identifying random noise over the entire set would also be useful):

(3) Noise probability = subNoise[1] + ... subNoise[N]

(4) Sub Signal probability = sub * 1/noise probability

Not much of a difference from the norm but the idea is to use the noise in all the subs to help better identify.

I'm thinking that a simple 3D FFT filter would probably help with (3).

[RichM63]

This stirs a memory. Years ago I was tasked with proof reading papers from the ENSSAT, check out the research and works of Kacem Chehdi in the field of noise recognition using statistical measures and adaptive filtering.

The physics and maths were completely beyond me, but the field rings a bell with your post. (eg. Noise reduction C0211923.pdf ) it might not be what you are looking for as it was over 10 years ago when I was working there.

[glappkaeft]

I've just had a bit of a brain storm. As most people throw away bad subs rather than use them. Up for a bit of a discussion...

 

Why not use bad sub noise to help statistically predict noise rather than throw them away, possibly allowing faint signal also to be identified in the good subs.

I think you are probably misunderstanding what signal and noise is. Signal is the number of electrons detected in the pixels; noise is the uncertainty of this electron count. The only thing you can know about the noise (in astrophotography where the noise is Poisson distributed) is that the expected magnitude of the noise is the square root of the signal. The main signals in an astrophoto is the bias (a constant level added to avoid zero-clipping the values in the A/D-converter), the dark current signal (electron leakage) and the photon signal (photons converted to electrons). These signals all have their own noise.

Unfortunately you can’t remove noise but adding signals will cause the noise to grow slower than the signal since the noise is random and partly cancels out (thus improving the Signal to Noise ratio - SNR). The same thing happens with the noise when you are subtracting signals except that since you are subtracting signal the SNR will deteriorate instead (less signal more noise).

I'm assuming you want to subtract the photon signal of the good image from the bad image. This will approximatively be a dark frame, except that the noise level will be much higher (you will have six sources of noise (2xBias + 2xDark + 2xPhoton) instead of two (Bias+Dark)) and all the differences between the frames will appears as artefacts. You will be much better off with regular darks.

Regards

Patrik