Square star "centroids"

Hi Marc,

Wide field data can have a very high density of stars and with sharp optics the debayer algorithm plays an important role here.

I have done quite some testing with widefield data and the development of the Airy Disc Debayer algorithm. In most cases the AAD algorithm gives the sharpest & roundest stars, less color edges around the stars and less chromatic noise in the background when compared to other debayer algorithms.

Some questions:

• First of all which debayer algortihm did you use in PI? I guess VNG?
• secondly, to be able to do a good comparision we need to normalize both linear stacks and apply an identical stretch on both stacks. Only then, we are able to say anything about the sharpness, roundness, color edges around the stars and the chromatic noise in the background. To evaluate visually it will help to significnatly boost saturation.
• compare the images when zoomed in 100% or further, otherwise the way the application draws the data on your screen pixels could mask the real differences on pixel scale. (this has to do with filter downscale operations when drawing a big image in s screen/panel with less resolution.

In your case, it could be the case that especially the smaller stars seem a bit blocky. I'll be happy to invetigate so please send me your 2 stacks and I'll have a look.

Edit: have received your stacks 😉

Possibly I could make another variant of AAD, but I will first investigate your stacks.

Cheers,

Mabula

Hi Mabula,

• First of all which debayer algortihm did you use in PI? I guess VNG?

Yes indeed, I used the VNG algoritme in PI.

In your case, it could be the case that especially the smaller stars seem a bit blocky. I'll be happy to invetigate so please send me your 2 stacks and I'll have a look.

Good to know, but I've seen this effect not only in my data. So a little investigation could be worthwhile. 

Regards,

Marc

Haven't looked yet at the stacks when normalised, but I will 😉

Don't you notice that VNG debayering actually makes your small stars square, only 45 degrees tilted?

Mabula

No, I can't see the tilt in the right PI stack compared to the left APP stack.

Hi Mabula - FWIW, I'm collecting some 50mm narrowband data with my mono QSI690.  I had noticed similar square shaped stars with APP versus rounder stars with Images Plus.  Of course, this is without any debayer algorithm and with an imaging train that produced extremely small stars.  I noticed this effect in my integrated stack - I didn't look at the individual subs, but I'm pretty convinced it was a result of the stacking algorithm.  Also, some of the stacking algorithms tended to produce dark halos - similar to what one sees when not applying sharpening.

Now, the details in the APP stack were sharper than the IP stack (again, similar to a high degree of sharpening).  I tried different pixel interpolation algorithms.  In general, for this data I found that Catmull-Rom spline tended to eliminate the dark halos and make for rounder stars.  But, that also meant some loss of sharpness.  If I recall correctly, Cubic B-spline seemed to work well, too.  But, the image was even softer.  The various Lanczos methods (Lanczos-3 being the default) was probably the best in terms of sharpness and worst in terms of square stars and halos.

For my comparisons, I was comparing the images using identical stretches.

I would appreciate any advice you might have as to what might make for best stacking of this type of mono data.

Scott

Hi Marc & Scott, thank you for your concern about the star shapes.

To get to the bottom of this we need to exclude all possibilities that could cause these squared star shapes that you report.

1. do you see a change when you use "no under overshoot" for Lanczos-3 data interpolation in 6) INTEGRATE? check this question of Donghun yesterday: tutorials-workflows/interpolation-artifacts If you don't enable this, then the Lanczos-3 data interpolation can cause ringing. This probably is what Scott is describing. Check these 2 wikipedia links as well in Donghun's question. I know that if you don't enable this, especially around small stars you will start to see lots of artefacts, so let me know if this setting improves things. Maybe I should set the "no under overshoot" to on by default. Donghun reported it worked excellent.
2. Have you checked how those stars look in APP's image viewer and compared them to how PI shows them? We need to exclude that the graphical screen drawing isn't a factor here.
3. Please disable outlier rejection in a new test. Outlier rejection could actually cause this, although I don't suspect this on both of your datasets, but to be sure, disable it and check the stars again.

One suggestion to Marc, don't use Linear Fit Clipping when you enjoy the benefits of LNC 😉 LNC actually is a solution for bad normalization, and bad normalization is the cause of the Linear Fit Clipping invention. LNC makes a linear fit clip filter redundant. (And the linear fit clip filter is much slower, I have never used it since I gave light to LNC 😉 )

In addition: since Marc reports this on DSLR RGB data, and Scott reports this on monochrome data, chances are the Adaptive Airy Disc is not the issue here. So maybe the cause is to be found elsewhere, for instance, maybe enabling "no under overshoot" is very crucial in very sharp data with very small stars... And that would be great to know. Both of you are having widefield data with similar focal lengths (both 50mm ?) and a very good and sharp objective as I recall from both of you ;-), (Zuiko and Zeiss ? )

Our aim should be to be able to use Lanczos-3 (which is simply the sharpest and best) without getting your square stars.

Could you run a few tests with your data with "no under overshoot" enabled and with outlier rejection disabled, so we can get to the bottom of this?

Cheers,

Mabula

Hi Mabula,

I can answer some questions. Enabling NUOS doesn't solve the square star issue, I tested this yesterday on a new data.

Regarding the APP image viewer vs PI, same issue, both visible.

I will test with suggestion 3 and your further suggestions.

Regards,

Marc

Hi Marc,

Thank you,

2 other possibilities to test with your RGB data

• try Bilinear debayer with your data and let me know how that looks
• don't debayer, but integrate using bayer drizzle, scale 1.0, top hat filter, droplet size 2

Let me know how that looks 😉

Mabula

This is a comparison (Canon 6Da) between (Monkeyhead nebula data set) PI (VNG) and APP (AAD).

Hi Mabula - I tried restacking my 50mm data with Lanczos 3 and No Under/Overshoot checked.  It resolved the ringing problem, but the stars still look pretty square to me.

I would say that No Under/Overshoot probably should be the default, unless there are some more common scenarios when it would create some other problem.  Ringing is definitely an issue most of us don't want in our stacks (it's too easy to add that later!)

Scott

Thank you Kees for contributing to this topic as well.

(In this case I clearly see square stars in PI (VNG) as well, rotated 45 degrees. But I should focus on APP off course)

I am waiting for Marc and Scott to give more feedback about running stacks with different settings to see what influences this.

Posted by: Scott Rosen

Hi Mabula - I tried restacking my 50mm data with Lanczos 3 and No Under/Overshoot checked.  It resolved the ringing problem, but the stars still look pretty square to me.

I would say that No Under/Overshoot probably should be the default, unless there are some more common scenarios when it would create some other problem.  Ringing is definitely an issue most of us don't want in our stacks (it's too easy to add that later!)

Scott

Okay, and still monochrome data right?

I think it's safe to have "no under/overshoot" on everywhere by default. I know that the way I implemented this, this hardly has any effect on the FWHM value of the stars. The stars will be a little bit bigger, in the order of 0.01-0.1 FWHM in pixels depending on the data.

Another idea to test, to see if the lanczos algorithm itself it the cause, try integrations with an increasing amount of frames, so 2,4,8,16,32,64 frame integrations and check the star shapes in all those integrations. It would also help if you could post 2 crops of some stars, one with LZ3 + NUOS and one with Cubic B-spline + NUOS.

Cheers,

Mabula

I'm not sure if the same interpolation algorithm is used but the stars on the APP version looks tighter in this case.

If the image is undersampled and well focused, isn't a star supposed to look square?

Donghun

Posted by: Dong Hun

I'm not sure if the same interpolation algorithm is used but the stars on the APP version looks tighter in this case.

If the image is undersampled and well focused, isn't a star supposed to look square?

Donghun

 

In the image of Kees and in the image of Marc at the start of this topic, AAD debayer is used, and besides data interpolation this could certainly have an effect as well.

But I agree, the stars in the APP stack are better and sharper in the case of Kees (what's in a name...), the VNG in PI even gives 45 degree rotated square stars which are bigger than 2pixels FWHM. Surely, I can't be the only one seeing this...?