Reducer Myth revisited

[ollypenrice]

12 minutes ago, vlaiv said:

In principle I agree with that statement, however it is missing important piece of information to make it complete.

What type of resampling are we talking about? There are different resampling algorithms with different properties.

For example - nearest neighbor resampling will make 0 improvement in SNR - it will leave SNR exactly the same. Binning as a form of resampling has very predictable improvement in SNR. Other forms of resampling will have SNR improvement somewhere in between (or even more due to pixel to pixel correlation).

Like I mentioned before, difference between binning and using equivalent focal reducer is in read noise contribution and covered FOV. Other forms of resampling (those that are worth using) will have slightly smaller SNR benefit compared to binning / equivalent reducer.

On the other hand upsampling behaves differently.

For most people that don't resample and have no clue about binning and such - focal reducer will provide real benefit in SNR, especially if they are oversampling at native focal length.

Yes, I'm not competent in all the ways of resampling because I don't do it, other than to compensate for JPEG artefacts when posting, but I really don't believe that there will be a serious difference between them. A slight difference, perhaps. We really need examples.

And, yes, if you are oversampling you might as well bin. (Another reason for not using one shot colour!) My graphic efforts did not allow for this distinction but I prepared them for use in a discussion about the optical side of imaging.

This might be a time to consider a few practical realities as well:

- Some CCDs bin not at all, or bin badly. Beware.

- Many experienced imagers do not find that Bin2 gives anything like a 4x reduction in exposure time in reality.

- Very fast optics will usually gain time only after they have cost the imager many nights of wasted time.

- Large pixel cameras are not available to the amateur with comparable sensitivity to small pixel cameras.

- Mysteries. Why do I greatly prefer processing my Atik 11000 data over data from an Atik 460 and a Moravian 8300? Because I'm not fighting the data, particularly on colour. But why not? I don't know. I just know that it isn't up for sale or replacement.

Olly

 

 

[ollypenrice]

1 hour ago, Adam J said:

You kinda looking at it the wrong way around. 

Lets say I have a 102mm F7 and I want to image medium galaxies ~M101. 

I can do one of two things. 

I can image at native focal length with an ASI 1600mm pro and get around 1.1" / pix.

Or I can use a reducer and get F5.6 and 1.4" / pix

The objects fits in the FOV in both cases. 

I would most likely chose to reduce the scope with a 0.8x reducer and go with 1.42 / pix. I would do that because I dont believe that with my skys and mount I can acheive 1.1" / pix. 

In this situation I would use the reducer because I wont loose real world detail in my image but I will gain some SNR. 

But here is the thing. I can most likely software re-sample the 1.1"/ pix image down to 1.4"/pix or even less in software and get a similar SNR improvement. Hence the reducer is not really adding much in this case. Its marginally better than software re-sampling. 

 

On the other hand. 

If with my camera sensor size I needed a 550mm focal length to fit my target into the FOV and I had a  80mm F7.5 then I would need to use my 0.85x reducer to fit the object into the frame. Hence software re-sampling is not going to help me and I must use the reducer. 

My other options would be to use a 100mm F5.5 scope. If I use the 100mm F5.5 scope then I will also fit my target in, but it will be at a faster f-ratio than the reduced 80mm F7.5 and so I will image in a shorter time. 

Or I could use a 70mm scope at F8. Thats probably not the best option but all of these give allow me to fit my target perfectly into my FOV. 

 

Lots of this come down to affordability. Its more affordable to own a 80mm F7.5 and reduce it than it is to own a 100mm F5.5 and operate it at native focal length.  By a factor of about 3x the cost.  

 

That help?

Adam 

 

I agree with all of this.

Olly

[Rodd]

3 minutes ago, ollypenrice said:

Mysteries. Why do I greatly prefer processing my Atik 11000 data over data from an Atik 460 and a Moravian 8300? Because I'm not fighting the data, particularly on colour. But why not? I don't know. I just know that it isn't up for sale or replacement.

I would hesitate to believe that you would get the same results with the FSQ and the 11000 sensor if reduced by .6.  Then again, the FOV of the 11000 sensor would make using the .6x reducer unnecessary.  OK--would it be reasonable or foolish for me to sell my cameras and possibly a scope to obtain an 11000 camera? 

[ollypenrice]

3 minutes ago, Rodd said:

I would hesitate to believe that you would get the same results with the FSQ and the 11000 sensor if reduced by .6.  Then again, the FOV of the 11000 sensor would make using the .6x reducer unnecessary.  OK--would it be reasonable or foolish for me to sell my cameras and possibly a scope to obtain an 11000 camera? 

Possibly not, because some of my preference may arise from the fact that I learned my processing on the Atik 4000 and 11000 which are very similar beasts. Maybe I've settled into habits which are closely related to these cameras. How would I know?

Olly

Edit. 0.6 reducer for me? No, 3.5"PP is my limit.

Edited October 18, 2019 by ollypenrice

[Adam J]

2 minutes ago, Rodd said:

I would hesitate to believe that you would get the same results with the FSQ and the 11000 sensor if reduced by .6.  Then again, the FOV of the 11000 sensor would make using the .6x reducer unnecessary.  OK--would it be reasonable or foolish for me to sell my cameras and possibly a scope to obtain an 11000 camera? 

yes because the QHY600 is coming out and that will bin in hardware up to the size of the 11000 pixels and have higher QE and lower read noise.

[Davey-T]

Regarding solar imaging using .5X focal reducer gives a smaller brighter image and the camera runs at 30 plus fps, without focal reducer you get a bigger darker image requiring fiddling with the settings to brighten it resulting in the camera running at 10fps.

Just saying 

Dave

[Rodd]

5 minutes ago, Adam J said:

yes because the QHY600 is coming out and that will bin in hardware up to the size of the 11000 pixels and have higher QE and lower read noise.

I hope it works as promised--we all know what happened in that respect with the ASi 1600

[Adam J]

2 minutes ago, Rodd said:

I hope it works as promised--we all know what happened in that respect with the ASi 1600

I dont have massive issues with my asi1600mm pro considering the cost is 25% of the QHY600.

But in honesty if you really want speed go with a duel rig mate.

Edited October 18, 2019 by Adam J

[Rodd]

3 minutes ago, Adam J said:

I dont have massive issues with my asi1600mm pro considering the cost is 25% of the QHY600.

Cost is more than the money--its the endless hours spent imaging and processing  an image ruined by microlensing.  To see an otherwise decent image of M13 ruined by microlensed 4th magnitude stars--to me that is a massive problem----for that target.  When it doesn't happen, your right, no problems.   Its every bit as frustrating as a miscollimated petzval when it happens

Rodd

PS--I think the price will come down

Edited October 18, 2019 by Rodd

[Davey-T]

4 minutes ago, Adam J said:

But in honesty if you really want speed go with a duel rig mate.

Do you want to have a duel with my dual rig Adam 

Dave

 

[Rodd]

2 minutes ago, Davey-T said:

Do you want to have a duel with my dual rig Adam 

Dave

 

2 scopes, 2 cameras, 2 computers---2 sets of cables.   Not easy to upgrade to--requires a new mount probably.  Not a very practical goal unless one starts out with that in mind.  Not many do

Rodd

 

Sorry--meant to quote Adam it seems.  Its all so confusing

Edited October 18, 2019 by Rodd

[Davey-T]

3 minutes ago, Rodd said:

2 scopes, 2 cameras, 2 computers---2 sets of cables.   Not easy to upgrade to--requires a new mount probably.  Not a very practical goal unless one starts out with that in mind.  Not many do

Rodd

Aiming for triple rig Rodd but some things don't want to co-exist in threes, my policy is to keep throwing money at it hoping it all works before I get too old 😂

ATB

Dave

PS: only one computer

Edited October 18, 2019 by Davey-T

[Rodd]

1 minute ago, Davey-T said:

Aiming for triple rig Rodd but some things don't want to co-exist in threes, my policy is to keep throwing money at it hoping it all works before I get too old 😂

ATB

Dave

PS: only one computer

2 dual rigs sounds better in that case--adds a potential of working on multiple targets.  But look, I have see images made with a modest refractor, a DSLR, taken from a Bortle 6 or wosre sky, using only  4 hours of data that absolutely trounce any image I can create.  I seriously do not think all the dual rigs in the world will change that

Rodd

[Davey-T]

I do sometimes have three mounts running on different targets but don't often get the conditions.

Dave

[Rodd]

1 minute ago, Davey-T said:

I do sometimes have three mounts running on different targets but don't often get the conditions.

Dave

Wow.  i would like to get another mount.

[ollypenrice]

1 hour ago, Adam J said:

yes because the QHY600 is coming out and that will bin in hardware up to the size of the 11000 pixels and have higher QE and lower read noise.

...which looks great, for sure, but it has to work for real, with no unpredicted difficulties not contained in the numbers. That's the perennial bugbear. Both of the other cameras I use throw up things which make processing harder. I wish I knew why...

Olly

[Adam J]

3 hours ago, Rodd said:

Cost is more than the money--its the endless hours spent imaging and processing  an image ruined by microlensing.  To see an otherwise decent image of M13 ruined by microlensed 4th magnitude stars--to me that is a massive problem----for that target.  When it doesn't happen, your right, no problems.   Its every bit as frustrating as a miscollimated petzval when it happens

Rodd

PS--I think the price will come down

You get micro lensing on M13!?!?!

[Rodd]

1 hour ago, Adam J said:

You get micro lensing on M13!?!?!

On the 2 bright stars near it. I am not at v CB imputes now so can’t provide link but it’d on my astrobin page

[jjosefsen]

Very interesting (and confusing thread) 🙂

How would you go about resampling (is that the correct term) to a certain resolution in PixInsight?

Example:

Let's say I have an image shot at 1"/px and I want to target 1.8"/px because I feel like that is what my skies allow. How would I do that?

Hope it's ok to ask in this discussion..

 

Edited October 18, 2019 by jjosefsen

[HunterHarling]

7 hours ago, ollypenrice said:

Both of the other cameras I use throw up things which make processing harder.

Olly-

Do you have some examples of these? Just curious.

[Rodd]

5 hours ago, jjosefsen said:

Very interesting (and confusing thread) 🙂

How would you go about resampling (is that the correct term) to a certain resolution in PixInsight?

Example:

Let's say I have an image shot at 1"/px and I want to target 1.8"/px because I feel like that is what my skies allow. How would I do that?

Hope it's ok to ask in this discussion..

 

There are 2 resampling tools. Integer resampling and resampling. I am under the impression that integer resampling is software binning, but they may both be. I resampling 2x2 binning by using the integer resampling tool set at 2. I do not know if it will do fractions. For that you may have to use resampling.  They both seem to do about the same. Others probably know more.  I have asked this on the PI forum the other day but have gotten no response yet
Rodd

[ollypenrice]

4 hours ago, HunterHarling said:

Olly-

Do you have some examples of these? Just curious.

The Atik 460 tends to have noise in the form of overly dark pixels in the background sky and its star colour is - how can I say - not reliable or convincing and needs working on. It is also hard to stretch up the background sky without blowing out the rest. I end up settling for a darker sky than I prefer. The Moravian 8300 produces a colour balance at parity of exposure which needs considerable adjustment in post processing and the colour data seems curiously weak. I haven't been using it for all that long so it may be my fault, but I've done darks and flats absolutely by the book several times. It's also possible that the filters are not as they should be but they are in a size of which I have no others to swap round.

Olly

[ollypenrice]

Adam and Vlaiv in particular, how about this caveat for the diagrams I posted?

These diagrams assume workable sampling rates for the camera in all cases. If the longer focal length introduces over-sampling it will add no resolution and the imager would benefit from the reducer since it will add speed without reducing resolution. An alternative would be to bin the data when oversampled.

I'm left wondering if we don't need a new unit. Arcseconds per pixel is fine for resolution but don't we really need something indicating flux per pixel? This might be indicated by something like square mm of aperture per pixel, no? 

Olly

 

Edited October 19, 2019 by ollypenrice

[vlaiv]

9 hours ago, jjosefsen said:

Very interesting (and confusing thread) 🙂

How would you go about resampling (is that the correct term) to a certain resolution in PixInsight?

Example:

Let's say I have an image shot at 1"/px and I want to target 1.8"/px because I feel like that is what my skies allow. How would I do that?

Hope it's ok to ask in this discussion..

 

You have couple of options there, not all are available in PI.

You can do fractional binning. That is whole new topic, and I'm just mentioning it as information.

Next thing to do would be simple resample. Depending on type of interpolation algorithm used for resampling you will get different results in SNR improvement (I'll expand on that with examples). Next thing that you could do is bin 2x2 to get to 2"/px and then upsample image back to 1.8"/px. This one will "cost" some of detail - but it's likely that most people won't be able to tell the difference. It will provide you with precise SNR improvement.

14 hours ago, ollypenrice said:

Yes, I'm not competent in all the ways of resampling because I don't do it, other than to compensate for JPEG artefacts when posting, but I really don't believe that there will be a serious difference between them. A slight difference, perhaps. We really need examples.

Here is one setup that will show effects of different resampling methods. It will be "artificial" / generated example, but because of that we will be able to get exact numbers.

I've created an image consisting out of two elements - a gaussian star profile with FWHM 2" and Gaussian noise with magnitude 1. Image is "sampled" at 0.5"/px. Here is image and measurements:

This image can be reduced to half its size ("sampling" at 1"/px) without loss of information - It is oversampled. I'm going to use different methods for reducing it and we can compare impact on SNR (measurement of noise - signal stays the same when resampling so any reduction in noise will be improvement in SNR) and also on any resolution loss - by examining how FWHM changes in reduced image.

In order, I will perform:

1. Nearest neighbor resampling (or rather just taking every other sample) - this approach should not have any impact on SNR and resolution what so ever

2. Regular binning 2x2 - it should reduce noise by factor of 2 and have small increase in FWHM due to pixel blur

3. Split / Sift binning - that is something I developed, so we can compare it to the others

4. Resampling with use of linear interpolation

5. Resampling with use of cubic interpolation

6. Cubic B-Spline

7. Cubic O-Moms

8. Quintic Spline

Here is a screen shot of first option:

As predicted, noise remains the same, and FWHM is also about the same (both vary because image is noisy) - now we are at 1"/px - so baseline FWHM is 2". Other measurements I'll just list instead of taking screen shots.

First SNR:

Results are in StdDev column. Baseline, and nearest neighbor have no change in SNR. Binning x2 and linear interpolation give SNR increase (predictably) by factor of x2. In this particular case linear interpolation is in fact bin x2 with half pixel shift (bin is average of two pixels, and because we are doing resizing by factor of x2 - linear interpolation is %50 of one pixel and 50% of other - which is the same as average).

Split/Sift bin gives the best results. We can go a bit deeper into that, but I've already gave outline explanation for this in thread on software binning that I gave link to in one of earlier posts. Other advanced resampling methods give poorer results. That is to be expected because advanced resampling methods are designed to do least alteration to image when resampling - that includes noise as well as data. In fact - most advanced resampling here gives SNR increase of only ~x1.2 vs split/sift bin that results in ~2.22 boost in SNR.

Let's now look at effects on resolution:

FWHMx	 2.0762 	FWHMy	 2.086
FWHMx	 2.0164 	FWHMy	 2.0235
FWHMx	 2.0762 	FWHMy	 2.086
FWHMx	 2.0007 	FWHMy	 2.0125
FWHMx	 1.9842 	FWHMy	 1.9988
FWHMx	 1.9798 	FWHMy	 1.9957
FWHMx	 1.9778 	FWHMy	 1.9946

Due to pixel blur, regular binning increases star FWHM by almost 4%  - 2.08" vs 2". Same is of course true for bilinear resampling. Split/Sift bin increases FWHM by about 1% (it's designed to circumvent pixel blur). More advanced resampling methods pretty much keep FWHM the same - they are designed so that they don't loose any information.

This little exercise also shows that for FWHM 2" stars, you don't loose anything when sampling at 1"/px - in fact, proper sampling for 2" FWHM image is about 1.25"/px (FWHM divided with about 1.6).

[Rodd]

5 minutes ago, vlaiv said:

You have couple of options there, not all are available in PI.

You can do fractional binning. That is whole new topic, and I'm just mentioning it as information.

Next thing to do would be simple resample. Depending on type of interpolation algorithm used for resampling you will get different results in SNR improvement (I'll expand on that with examples). Next thing that you could do is bin 2x2 to get to 2"/px and then upsample image back to 1.8"/px. This one will "cost" some of detail - but it's likely that most people won't be able to tell the difference. It will provide you with precise SNR improvement.

Here is one setup that will show effects of different resampling methods. It will be "artificial" / generated example, but because of that we will be able to get exact numbers.

I've created an image consisting out of two elements - a gaussian star profile with FWHM 2" and Gaussian noise with magnitude 1. Image is "sampled" at 0.5"/px. Here is image and measurements:

This image can be reduced to half its size ("sampling" at 1"/px) without loss of information - It is oversampled. I'm going to use different methods for reducing it and we can compare impact on SNR (measurement of noise - signal stays the same when resampling so any reduction in noise will be improvement in SNR) and also on any resolution loss - by examining how FWHM changes in reduced image.

In order, I will perform:

1. Nearest neighbor resampling (or rather just taking every other sample) - this approach should not have any impact on SNR and resolution what so ever

2. Regular binning 2x2 - it should reduce noise by factor of 2 and have small increase in FWHM due to pixel blur

3. Split / Sift binning - that is something I developed, so we can compare it to the others

4. Resampling with use of linear interpolation

5. Resampling with use of cubic interpolation

6. Cubic B-Spline

7. Cubic O-Moms

8. Quintic Spline

Here is a screen shot of first option:

As predicted, noise remains the same, and FWHM is also about the same (both vary because image is noisy) - now we are at 1"/px - so baseline FWHM is 2". Other measurements I'll just list instead of taking screen shots.

First SNR:

Results are in StdDev column. Baseline, and nearest neighbor have no change in SNR. Binning x2 and linear interpolation give SNR increase (predictably) by factor of x2. In this particular case linear interpolation is in fact bin x2 with half pixel shift (bin is average of two pixels, and because we are doing resizing by factor of x2 - linear interpolation is %50 of one pixel and 50% of other - which is the same as average).

Split/Sift bin gives the best results. We can go a bit deeper into that, but I've already gave outline explanation for this in thread on software binning that I gave link to in one of earlier posts. Other advanced resampling methods give poorer results. That is to be expected because advanced resampling methods are designed to do least alteration to image when resampling - that includes noise as well as data. In fact - most advanced resampling here gives SNR increase of only ~x1.2 vs split/sift bin that results in ~2.22 boost in SNR.

Let's now look at effects on resolution:

FWHMx	 2.0762 	FWHMy	 2.086
FWHMx	 2.0164 	FWHMy	 2.0235
FWHMx	 2.0762 	FWHMy	 2.086
FWHMx	 2.0007 	FWHMy	 2.0125
FWHMx	 1.9842 	FWHMy	 1.9988
FWHMx	 1.9798 	FWHMy	 1.9957
FWHMx	 1.9778 	FWHMy	 1.9946

Due to pixel blur, regular binning increases star FWHM by almost 4%  - 2.08" vs 2". Same is of course true for bilinear resampling. Split/Sift bin increases FWHM by about 1% (it's designed to circumvent pixel blur). More advanced resampling methods pretty much keep FWHM the same - they are designed so that they don't loose any information.

This little exercise also shows that for FWHM 2" stars, you don't loose anything when sampling at 1"/px - in fact, proper sampling for 2" FWHM image is about 1.25"/px (FWHM divided with about 1.6).

I think the question was more how to do it in PI but great info. The question remains, how does this translate to PI