Decon for non-simple star distortions?

[aworonow]

I understand the advantages of modeling stars using certain Point-spread distribution functions, an then utilizing the models for the decon. This is a "robust" way of treating the observations. However, the models may not be appropriate in all instances, and being restricted to a set  of non-applicable models may lead to poor decon. For instance, I have a critical image where something went temporarily wrong in the  tracking, leading to an image in which the stars have two more-or-less Gaussian shapes (unequal, it appears) overlapping--kind of like a dumbbell of sorts. I think, if I could simply produce a PSF that had this shape, I might decon the image to improve its appearance and improve its resolution in the targeted nebula. Maybe simply fitting some smoothed terrain to 1 or several dumbbells might work? Anyway, this kind of distortion is beyond current repair, I think, but would be a valuable and unique addition to the arsenal of PI tools.
BTW, I have tried the blur decon, and it does round-out the stars, but it also leaves a bright band across the middle of the dumbbell, perpendicular to the axis of the dumbbell.

Alex

[pfile]

since Deconvolution supports arbitrary PSF images, could you try to construct the PSF image in some other way than using DynamicPSF? maybe crop out one of these double stars from the image and the blur it?

rob

[aworonow]

Good idea. I did not realize that the PSF image could be edited and still be viable. Maybe a "drawing" using some paint tool on a bitmap would work too?

[dnault42]

I have a script I wrote some time ago that does exactly this.  It uses Juan's code to detect stars then take a capture of all the stars and scales them to match before stacking them into a single PSF.  You can download the script from here.

To install the script copy it to your <PixInsightInstallDir>/src/scripts then in PixInsight go to SCRIPT->’Feature Scripts…’ and click the ‘Add’ button.  It will ask you for the directory, point it to your <PixInsightInstallDir>/src/scripts folder and hit OK.  It should find the script and make it available in the SCRIPT menu.  It will show up under SCRIPT->Render->dnaGeneratePSF.

It will also generate a synthetic star field showing which stars it is using and what it thinks their synthetic PSF looks like.  If you see a lot of oddness in that image your result probably won't be very good and you should change some of the star detection parameters.

I always hesitate to release one of my scripts to the public because, while the work for me, I have no idea how well they will work for others and I'm not really a developer who can go off and fix a lot of bugs and make the code more robust.  So I put this out there in the warning that the results may not be what you want.  On occassion I've had the size of the PSF render box be way to large and it has to be cropped back down and re-normalized manually.  Sometimes it has captured galaxies or other structures or incorrectly determined the PSF of a star as part of it's evaluation and generates incorrect scaling of the pixel data in the final PSF.  Those cases tend to be rare in my data but I only have my data to use as reference most of the time.

Regards,
David Ault

[KuriousGeorge]

Hi folks,

I'm wondering why this thread stalled so quickly? There's seems to be a huge potential here assuming Regularized Richardson-Lucy (or some other) can form a decent mathematical model around an actual PSF.

The attached PSF is from an optical system having a mirror issue. Although Deconvolution sharpens the stars dramatically with the "actual" PSF, it doesn't shape them. I'm wondering if this is because the algorithm is trying to approximate the PSF with a traditional star model rather than using the PSF as is (i.e., a more advanced fitting model)? I notice DynamicPSF always causes a "smooth" PSF which is not necessarily representative of the data.

In this case the coma issue was fairly uniform throughout the frame, especially near the subject which is a small percentage of the 4K x 4K frame. In the more generic case, the image can be divided into smaller segments with a PSF calculated for each segment.

Is this as a huge opportunity for advancing our deconvolution quality or am I missing something?

Looking forward to your feedback!