Impact of Deconvolution on LRGB Images

[dgbarar]

Hi All,

I have been working with Pixinsight since late November 2011.  This program has helped me learn much about astro image processing.  But it has made me far more critical of of my work than I have ever been in the past.  I believe the equipment that I am working with is good: Paramount ME  and a  post-2000 AP 155 EDFS.  The ST10 XME camera I use might be the weakest link as the covered slipped 3200ME CCD can be prone to reflections on bright stars.

One of the things that I have noticed is that the use of deconvolution results in colored halos on stars in LRGBs.  Attached is an M15 example "Deconvolution Impact on LRGB.jpg". In assembling the LRGBs, I attempted

1) Stretch to about the same on both the RGB and L images using masked stretch to a median of 0.30 and a final histogram adjustment to about 0.09.  There is some variation in the median pixel of the final RGB as a result of the color balance.
2) Deconvolve the channels the same way with respect number of iterations and dark side deringing.  The PSF, star masks, and background masks were different for each channel.

The example consists of three columns.

1) Column 1 is with no deconvolution.  L channel stars are about the same size as the RGB image.  Halos around the stars are minimal but the stars are larger than I would like them to be.  Since this was an image of M15, I wanted to have more granularitity in the image.
2) Column 2 is with deconvolution on the L channel only.   The L channel stars are significantly smaller.  But when combined with an RGB where the channels have not been deconvolved there are colored halos--blue being the most significant.
3) Column 3 has all the channels deconvolved.  I thought that by getting the RGB composite to have smaller star profiles through deconvolution on each of the R, G, B images that I would end up with less colored halos.  This did not seem to make much difference as the blue halos were still still there and about the same size.

I have also attached an image, Deconvolution by Channel.jpg, comparing the impact of deconvolution on the individual channels thinking that there may be a difference in star size reduction as a result of deconvolution.  There does not appear to be a difference in star sizes between the individual channels after deconvolution.

So I am at a bit of loss on how to correct the blue halos that result from deconvolving components in an LRGB.  I believe as though I am missing an important processing step.  Has anyone run into this problem before?  How did you solve it?

Cheers,

Don

[pfile]

i have seen something like this, but my solution was to compute the PSF for each of R, G, B independently and deconvolve each channel separately. it sounds like you have already done this.

my problem was a little different - using an OSC with refractive optics, my R channel was always slightly out of focus with respect to the G and B. your camera is mono so in theory this should not happen as long as focus was proper on each filter.

what is the average FHWM in each channel before deconvolution? i suppose that if they happen to be different, even though you have tuned the deringing and PSF for each channel, that the final FHWM in each channel will end up being different, thus generating halos when they are combined.

case #2 seems to indicate that the FHWM is already different - the colored halos can only come from the RGB data and it has not been deconvolved.

perhaps the answer is to only deconvolve the color channel with fat stars to try to bring them into line with the other two, and concentrate on the L for the "true" deconvolution...?

[dgbarar]

what is the average FHWM in each channel before deconvolution?

At your suggestion I used PI to calculate FWHM before and after deconvolution.  Follows are the results:

           Before Deconvolution       After Deconvolution   
Channel   FWHM X   FWHM Y      FWHM X   FWHM Y
L           2.00           1.87               0.69   0.63
R           2.06    1.90                    0.70    0.67
G           2.10           2.08               0.77   0.75
B           2.27    2.22               0.72   0.70

I expected that the FWHM of the blue channel before deconvolution was higher but not by much. However, after deconvolution all of the channels are about the same with the G channel being the greatest  So I am still puzzled by what I am doing in processing that results in the blue halos?  One thought I have is that the application of SCNR to removes excess green leaving blue halos.  But this is just a guess.

[Alejandro Tombolini]

Hi,
if i understad well you are using deconvolution after streching, and it must be applied before

[dgbarar]

Hi Alejandro,

As always,  deconvolution was performed on linear images before any stretching occured.  After deconvolution I performed histogram adjustments and cropped out small sections of the various images to show the impact.

Don

[pfile]

hm, the FHWM is pretty comparable before and after. the blue channel does look a little 'fatter' though.

i guess without playing with the data i'm out of ideas. one thing is, is it reasonable to expect such a drastic change after deconvolution? from 2 px to less than 1 seems like a lot.

[dgbarar]

Hi,

I too was also surprised by the dramatic decrease in FWHM following deconvolution.  I hope I determined FWHM correctly.  What I did was open dynamic PSF, select 20 stars (tried to select the same stars but it is difficult to do), and then selected the summation button to get the aggregate statistics including FWHM.

I would be delighted to upload the L, R, G, B channels for someone else to process.  But it looks like I am limited to 512 KB.

Don

[dgbarar]

Hi Again,

I have attempted to dig into this more.  I believe I understand what is happening but just don't know how to stop it from occurring while processing.  Maybe one of you can help me figure it out.

Attached is strip of images that shows a small section of the same region around M15 (Before After Deconvolution.jpg).  The top image is an LRGB with all channels deconvolved as previously discussed in this thread.  Below this small section of the LRGB is the same region for each of the channels before and after deconvolution.  I determined the FWHM for each of the channels on the same star in the upper most right that has the most severe blue halo.  The results are as follows:

                Before Deconvolution       After Deconvolution
Channel   FWHM X   FWHM Y      FWHM X   FWHM Y
     L         1.85         1.71                 0.82      0.74
     R         2.05         2.00                 0.82         0.78
     G         2.12         2.08                 0.89         0.87
     B         2.26         2.21                 0.80         0.79

As before, the deconvolved FWHM for that star are about the same for all of the channels with green being the largest.

Please look at the image set showing the before/after deconvolution.  The L, R, and G channels all deconvolve about the same with the star nicely folding onto itself to form a smaller dot.

The blue channel deconvolution appears different.  There still is a major drop in star size as with the other channels that is consistent with the smaller FWHM.  However, with the blue channel after deconvolution there is a halo left around the star that is about the same size as the blue halo seen in the LRGB.

My thinking is that I need to treat deconvolution of the blue channel differently to eliminate this halo.  But I do not know how to do it.  Is this an instance for bright side protection?

If someone could help me with this I would be most grateful as the quality of my images would significantly improve.

Cheers,

Don

[Alejandro Tombolini]

Hi Don,
You can apply deconvolution to each channel. In blue channel try decreasing deringing for the actual values or move it convenient for the halo disappear.
See this processing example where is use deconvolution with diferent parameters for each channel.
Also see this where Juan explain how to determinate the correct PSF.
hope this helps
Saludos. Alejandro.

[dgbarar]

Hi Alejandro,

The example I provided is with deringing set at 0.005.  At your suggestion I performed the following tests:

Dering Parameter       Result
0.005                          Halo around the star
0.004                          Bright center gets smaller.  Halo around the star does not change.
0.003                          Bright center gets smaller.  Halo around the star does not change.
0.002                          Bright center gets smaller.  Halo around the star does not change.
0.001                          Bright center gets smaller.  Dark ring between bright center and halo.  Halo around the star does not change.

It looks as though reducing the dering parameter further does not help.  It results in a tighter central core, but does not reduce the size of the halo.  If I reduce the dering parameter to much, I get a dark ring around the central bright core, but the light halo around the star persists and does not change diameter.

I also thought I would try the bright sides parameter.  The highest I could go was 0.04 but this doubled the FWHM to 1.6 and the halo did not go away.  Bright side parameters higher than 0.04 resulted in accumulated divergence errors and general messing up of the image.

Any other thoughts?

Cheers,

Don

[dgbarar]

Hi Alejandro,

Your last eMail where you made the suggestion about the correct PSF gave me a thought.  There can be no better PSF for a star than the star itself.  I generated a PSF based on the blue star with the halo and performed another deconvolution.  This made no difference.  The star reduced in size as before and a blue halo was the same.  I do not think this is a PSF issue.

Don

[Alejandro Tombolini]

Hi Don, maybe you can upload in endor u other site a crop of integrated L and RGB to make a try.
You can also try "StarHaloReducer" scrip in blue channel, but as I do not see the halo before deconvolution I think that it could be solve adjusting deconvolution parameters. 
Saludos. Alejandro.

[dgbarar]

Hi Alejandro,

Not sure I understood your first request.  StarHaloReducer script is not an option.  If all I had was a couple of stars that might be OK.  But in this case I have many stars that deconvolved like this one.

Don

[pfile]

endor is the pixinsight fileserver where you can upload large files. you need to register for an account with juan.

[dgbarar]

Hi,

I did a search on endor to determine how to make a request for an account.  Did not find anything accept this thread.  How do I make this request to Juan?

BTW.  I am starting to make some progress on this issue by going the Parametric PSF approach.  Not quite there yet.  I know what the parameters mean.  Just not certain how to optimize them to get the desired result.

Don