Should we be using Binning in light of AI noise reduction

[Adam J]

My basic theory is that when used purely for the purpose of noise reduction the use of Binning is made redundant by the advent of modern AI noise reduction techniques and in some cases my be detrimental to image quality. 

In effect AI noise reduction is making best use of available information contained within the raw image to reduce noise dynamically across the image dependent on local SNR. By binning the image you are simply depriving the AI algorithm of information and potentially missing out on detail in areas of already high SNR if whole image binning is applied. 

In the case of a oversampled image I would always now apply AI noise reduction and then as a final step resample the image to present it at an appropriate image scale. 

Thoughts? 

Adam

[andrew s]

I have not studied any of the AI noise reduction methods but your proposal, at face value, makes sense. 

All image processing loses information.  However, it allows us to modify an image so we can visually see the "features " in the image we want to bring out.

At a practical level it would be easy to do both (AI then bin or bin then AI) and compare the result.

Regards Andrew 

Edited May 2 by andrew s

[ollypenrice]

This is an interesting (and rather worrying) idea. Why worrying? Because when you make 40 panel mosaics you'd need an almighty computer to work at full size. 

But is binning used purely for the purpose of noise reduction? The idea is to get more signal per pixel and bigger 'effective pixels' do that.

Since it's wet outside I think I'd better get experimenting...

Olly

[gajjer]

1 minute ago, ollypenrice said:

But is binning used purely for the purpose of noise reduction? The idea is to get more signal per pixel and bigger 'effective pixels' do that.

That was my thought. I always use binning to get a higher gain but at the cost of resolution. But then I'm no expert.

cheers

gaj

[vlaiv]

3 hours ago, Adam J said:

and in some cases my be detrimental to image quality. 

Can you explain this?

When is binning detrimental to image quality?

[Adam J]

29 minutes ago, vlaiv said:

Can you explain this?

When is binning detrimental to image quality?

yes, it's detrimental when you are undersampled and you loose detail in bright areas to gain SNR in faint areas of the image. Something that occurs in wide field imaging quite allot. The point being in wide field you are almost never limited by seeing. 

My example is that when I processed my M45 wide image at 180mm focal length, the surrounding dust looks better if you bin the image as you might expect. But in binning the whole image you loose detail in the core of M45. 

Yes it would be possible to bin an image, resample back up to original level and then mask the brighter areas of the unbinned image back in but the AI noise reduction seems to achieve this with better granularity and minimal effort. 

[Adam J]

53 minutes ago, gajjer said:

That was my thought. I always use binning to get a higher gain but at the cost of resolution. But then I'm no expert.

cheers

gaj

Yes that's correct, but it's a blunt tool applied across the whole image, even the brighter areas. 

[vlaiv]

4 minutes ago, Adam J said:

yes, it's detrimental when you are undersampled and you loose detail in bright areas to gain SNR in faint areas of the image. Something that occurs in wide field imaging quite allot. The point being in wide field you are almost never limited by seeing. 

My example is that when I processed my M45 wide image at 180mm focal length, the surrounding dust looks better if you bin the image as you might expect. But in binning the whole image you loose detail in the core of M45. 

Yes it would be possible to bin an image, resample back up to original level and then mask the brighter areas of the unbinned image back in but the AI noise reduction seems to achieve this with better granularity and minimal effort. 

We seem to be having different concepts of what the binning is for.

I see it as a data gathering step - integral into data reduction stage, rather than final processing tool.

In my view - if you choose to bin your data - there is no sense in up sampling it. You don't up sample your data from regular pixels either, right?

You choose to bin your data because you'll be happy with final sampling rate that binned pixels give you - as if you were simply using larger pixels in the first place.

If you leave your image binned - it won't be any different than shooting "natively" at that sampling rate. It can loose detail versus properly sampled image - but that is not due to binning, same would happen if you compared two images taken with regular pixels - ones being bigger and under sampling and ones being smaller and properly sampling. Under sampled data will not show the same level of detail - simply because it's under sampled (not because it was binned - because in this case it was not).

 

[Adam J]

42 minutes ago, vlaiv said:

We seem to be having different concepts of what the binning is for.

I see it as a data gathering step - integral into data reduction stage, rather than final processing tool.

In my view - if you choose to bin your data - there is no sense in up sampling it. You don't up sample your data from regular pixels either, right?

You choose to bin your data because you'll be happy with final sampling rate that binned pixels give you - as if you were simply using larger pixels in the first place.

If you leave your image binned - it won't be any different than shooting "natively" at that sampling rate. It can loose detail versus properly sampled image - but that is not due to binning, same would happen if you compared two images taken with regular pixels - ones being bigger and under sampling and ones being smaller and properly sampling. Under sampled data will not show the same level of detail - simply because it's under sampled (not because it was binned - because in this case it was not).

 

I would say there are two reasons to bin, 

1) You want to improve SNR by effectively making the pixels bigger. 

2) You are oversampled, likely due to seeing and you want to present your image at a scale at which it appears sharp to the eye as opposed to soft. 

In the case of two I would argue that the AI noise reduction will provide a better result if used before you software bin and in this case I would apply noise reduction as a first step following stacking then resample. In the case of one I would just let the AI noise reduction do it's thing if the image is critically sampled or undersampled. If oversampled see case two. 

Adam

 

Edited May 2 by Adam J

[vlaiv]

6 minutes ago, Adam J said:

I would apply noise reduction as a first step following stacking then resample.

I would advise against this as stacking would not make sense after you alter noise distribution.

In any case - why don't you simply try it out on existing data you have? Make comparison between two approaches.

Maybe best way to do it would be to create "split screen" type of image. Register both stacks against the same sub so they are "compatible" - prepare the data one way and the other and compose final image prior to processing out of two halves - left and right copied from first and second method.

That way final processing will treat both datasets the same as you'll be doing it on single image.

Alternatively - if you can't make it work that way - just do regular comparison - do full process one way and then the other.

I'd be very happy to see the results.

[Adam J]

Just now, vlaiv said:

I would advise against this as stacking would not make sense after you alter noise distribution.

In any case - why don't you simply try it out on existing data you have? Make comparison between two approaches.

Maybe best way to do it would be to create "split screen" type of image. Register both stacks against the same sub so they are "compatible" - prepare the data one way and the other and compose final image prior to processing out of two halves - left and right copied from first and second method.

That way final processing will treat both datasets the same as you'll be doing it on single image.

Alternatively - if you can't make it work that way - just do regular comparison - do full process one way and then the other.

I'd be very happy to see the results.

I didn't say that though, I said I would apply it after stacking. 

I have tried it and the thread is to ask if others have observed the same thing. 

Adam

[vlaiv]

2 minutes ago, Adam J said:

I didn't say that though, I said I would apply it after stacking. 

Sorry, my bad, did not understand that correctly.

2 minutes ago, Adam J said:

I have tried it and the thread is to ask if others have observed the same thing. 

Have you posted the comparison?

[Adam J]

49 minutes ago, vlaiv said:

Sorry, my bad, did not understand that correctly.

Have you posted the comparison?

no I am on my phone, would need to do it later. 

[gajjer]

To clarify things for my simple mind. Are you saying that you take an image with binning at 1x1 and then, in processing it, you choose to bin 2x2 ( or whatever ).

I was thinking that you were binning at the time of taking.

cheers

gaj

[vlaiv]

1 minute ago, gajjer said:

To clarify things for my simple mind. Are you saying that you take an image with binning at 1x1 and then, in processing it, you choose to bin 2x2 ( or whatever ).

I was thinking that you were binning at the time of taking.

cheers

gaj

I personally don't mind either - or in fact prefer binning at data reduction stage - because there is technique to do it that is slightly better than using "firmware" binning.

However, when I say binning at data reduction stage - it is performed after calibration and before registration and stacking. It can be thought of as part of calibration stage.

[gorann]

4 hours ago, gajjer said:

To clarify things for my simple mind. Are you saying that you take an image with binning at 1x1 and then, in processing it, you choose to bin 2x2 ( or whatever ).

I was thinking that you were binning at the time of taking.

cheers

gaj

Interesting thread that I will follow! Regarding binning Gaj, it is with modern CMOS cameras usually done after taking the image (not during imaging as with your CCD 383L) since CMOS cannot be hardware binned. So for us CMOS users we normally collect unbinned subs since we can bin them later in software and the only reason to tell the camera to bin the data would be to save on hard drive space.

[gajjer]

21 hours ago, gorann said:

Regarding binning Gaj, it is with modern CMOS cameras usually done after taking the image (not during imaging as with your CCD 383L) since CMOS cannot be hardware binned. So for us CMOS users we normally collect unbinned subs since we can bin them later in software and the only reason to tell the camera to bin the data would be to save on hard drive space.

Thanks for that.

cheers

gaj