PixInsight Forum (historical)

PixInsight => General => Off-topic => Topic started by: nmontec on 2012 May 08 07:30:26

Title: Learning deconvolution
Post by: nmontec on 2012 May 08 07:30:26

I know the many times spelled mantra "you should not use deconvolution on non linear images".
But yesterday I tried to deconvolve an image of mine and with the linear image I was getting a lot of artifacts, while I was getting exactly what I wanted (star tightening) with the stretched one.
I asked myself why, and the only answer I got is: "deconvolution is very sensitive to SNR, therefore I get artifacts becaue in the linear image the SNR is lower then in the stretched one."
Is it this assumption correct ? If so there is an exception to the mantra, otherwise I might not be working with the right parameters...

Title: Re: Learning deconvolution
Post by: sreilly on 2012 May 09 00:49:20

It might help to describe the artifacts and/or post an image showing these along with the steps you used. To use Decon I first take the linear image and use Histogram Stretch. With that result I then make my star mask, saving the mask and throwing away the stretched image. Then I open the saved linear image again and apply the star mask and then use Decon. I also make sure that I have Ringing checked and the star mask listed in that "Local Support" location. See the attached screen shot.

Steve

Title: Re: Learning deconvolution
Post by: Raymond on 2012 May 09 04:03:55

For learning deconvolution, this tutorial helped me a lot:
http://pixinsight.com/forum/index.php?topic=4048.0 (http://pixinsight.com/forum/index.php?topic=4048.0)

Title: Re: Learning deconvolution
Post by: nmontec on 2012 May 09 06:49:02

Thanks Steve and Raymond.

@Steve: I do the same thing, and the result is attached (starmask and after deconvolution).

@Raymond: I read that and didn't help me on understanding how to proceed.  :(

Mask is nicer then the real image! :D

Title: Re: Learning deconvolution
Post by: cs_pixinsight on 2012 May 09 12:12:40

nmontec, with deconvolution on a linear debayered image you need to assign a linear RGB working space to the image prior to running deconvolution.  Use RGBWorkingSpace and set R=G=B=1 and Gamma=1 and drag the process triangle over the linear image.

Now you can proceed with the deconvolution steps.  Once you perform a non-linear stretch you need to change the RGB working space back to a nonlinear one for the rest of your processing steps.

Hope this helps,
Craig

Title: Re: Learning deconvolution
Post by: sreilly on 2012 May 09 12:53:25

I did not know this was with a DSLR image data. There are steps you need to take before processing the image and I'll have to step back as I have no idea what they are. I use a mono camera and filters.

Title: Re: Learning deconvolution
Post by: nmontec on 2012 May 09 13:49:12

In fact it is not a DSLR image...

Title: Re: Learning deconvolution
Post by: sreilly on 2012 May 09 14:11:32

Well maybe better clarification is needed. Just what has been done to the image in question from start to the image you are trying to use decon on.  And maybe the camera you are using.

Steve

Title: Re: Learning deconvolution
Post by: nmontec on 2012 May 09 15:20:23

A quick update. I managed to have smaller stars with almost no artifact.
The trick was:
- not using a global star mask
- using the star mask for the local deringing
- increasing the stddev to 4
- Inserting aspect ration and rotation as suggested by Manuel in his tutorial instead of relying on the calculated PSF

Soon I'll post the result in the appropriate section. Thanks a lot for your comments.