Anyone have experience with deconvolution using real PSFs?

[darkownt]

I have been doing some experiments, synthetic and real world, using a method to extract PSFs for my optics using various tools of PixInsight.  Of note is the fact that some of these PSFs are asymmetric (non eliptical off centered shapes), and have some form of bokeh of the kind where the center is darker than the outer edge.

I'm noticing that when working with actual PSFs that the Weiner and Constrained least squares deconvolution work much much better than iterative R-L or VanCittert deconvolution, which may be due to the nature of the actual PSFs and the assumptions made in the algorithms.  I am quite happy with using the image restoration deconvolution since the results are quite nice, BUT if there is the possibility for using iterative deconvolution to advantage I would like to be able to do so.

I was wondering if anyone has any tips for:

1.  Settings for iterative deconvolution which are more effective when using actual PSFs which are not gaussian and generally assymetric.

2.  Massaging real world PSFs to work better with either the restroration or deconvolution processes.  For example, if anyone else has found that real world PSFs need to be gamma adjusted slightly to sharpen or widen the function, or if it should be convoluted with a widening kernal (softened) or deconvoluted (sharpened) before use...

I have attached a simulated blurred image, the extracted PSF and a result using restoration.  For reference the image file from which the test was generated is also included.  It has some artifacting which is purposely included to see how well they are reproduced.

Any feedback would greatly be appreciated!

Colin

PS:  A synthetic PSF was applied to generate the simulated blurred image.  The file attached is the resulting PSF extracted from the blurred image which closely matches what was applied.

[mmirot]

How are extracting a PSF from this image?

Max

[Carlos Milovic]

Hi Colin

I agree with you. With some kernels the Wiener filter works better than RL (in most cases, RL is also better than Van Cittern). Also, it is faster.
I have not have too good results with "real" PSF, like using stars directly. I would say that if you may have a mean to simulate the data matrix, then go that way instead of using real data. If the latter is the only alternative, then I would process it with wavelets, to delete the first layer, or perform some noise reduction in the first 2-3 layers.

[darkownt]

Hi Colin

I agree with you. With some kernels the Wiener filter works better than RL (in most cases, RL is also better than Van Cittern). Also, it is faster.
I have not have too good results with "real" PSF, like using stars directly. I would say that if you may have a mean to simulate the data matrix, then go that

way instead of using real data. If the latter is the only alternative, then I would process it with wavelets, to delete the first layer, or perform some noise

reduction in the first 2-3 layers.

Hi Carlos:

Thanks for confirming my suspicions. 

FYI In my simulated/synthetic tests I convolve a reference image using a test PSF matrix to generate the blurred image.  In this sense, my tests do not suffer

problems of intenity dependent bloom (overloaded CCDs) or intensity saturation of pixels etc. due to real world dynamic range limits of CCDs  (See Question

below)  I have confirmed that the extracted PSF is the ACTUAL PSF that was applied.  Using the mathematically correct PSF ... as you say, Weiner and

Constrained Least Squares seems to do much better than RL or VanCittert, and I would guess that it is because of the mathematical form of the partiular PSF. 

I would guess using a gaussian PSF, R-L would do a fine job.  I will keep experimenting with the restoration filter and trying to massage the ACTUAL PSF to

compensate for the algorithms...

Questions:  I am still learning about astronomical photography so please excuse my ignorance. 

1) First I am curious when taking astronomical photographs which one of the following tends to be more true given the limitations and capabilities of CCDs,

for any given digital image of a star whose actual visual size (angle subtended... etc.) is less than one pixel (if the optics were perfect)

A) the CCD dynamic range is wide enough such that every star in the digital image is a "point source" spread --only-- by the actual spreading function caused

by the atmosphere, and the optics between the actual star and the CCD, and hence is the same function regardless of the brighness of each star in the image.

B) the CCD is such that the stars in the digital image are point sources spread by a function which depends on blooming, intensity saturation limits etc. of

the CDD such that each star effectively has a different spread function which depends on the brightness of that star


2)   Is it possible using a 16 bit CCD, to take multiple images at different exposures, and then remove intensity saturated pixels from each picture, and

recombine them with proper intensity scaling in order to produce a proper digital image like A) discussed above?   Would such a tool for recombinging images

be useful?


cheers
Colin

PS I'm trying to figure out how to adapt my PSF extraxtion methods to astronomical images

[Carlos Milovic]

Hi Colin

1.- A
It is assumed that the whole image has been convolved by the same function. This, of course, is a very simplistic approach, but works quite well in most of cases. Intensity dependent functions, as far as I know, are not used, since the FWHM is pretty much the same along stars of different brightness. By the other hand, spatially dependent convolutions are much more common, due the optics mainly. We wanted to implement such deconvolutions since a while, but there has been other priorities.

Oh, btw, blooming is ignored here.

2.- Yes. Such "field" is called high dynamic range processing. For this particular task we have a script called HDRComposition. It will also be released in short term as a processing module, redesigned.

[darkownt]

Hi Carlos:

Thanks for the response!

So now I'm looking for an accurate starfield image RAW or 16bit TIFF taken with a telescope which has no oversaturation, in native linear format so that all the PSF data is still in there and uncomprimised.  The PSF of the optics at any one point (center for instance) does not and cannot change based on the brightness of the star there, so I am very certain overexposure would be bad for my process. 

Such a high quality image is what I need to extract the PSF...  no any way could get my hands on such a picture to experiment?

 

cheers
Colin