Steps, strategies and FAQ for successful deconvolution on deep sky images with Pixinsight
Deconvolution is one of those complicated steps in image processing. It's quite difficult to get the correct PSF (Point Spread Function) for a given image. As you can see in this example of a "correct" deconvolution, the stars are tighter, and the resolution is a bit better.
This guide will try to cover the basic steps and the most frequent problems in order to improve your skills. There is always more room to improve, so don't take this as your bible. As usual, experiment by yourself, and find the best way to process in your own style.
The examples are generated with a 3 hour poor seeing capture of M51.
If you have any comments or questions, please contact me!.
These are the basic steps:
Deconvolution is done only on linear images!
1) Deringing mask creation Create a mask for the stars. It has to protect only the brightest ones, but it has to protect them well. Don't make it strong enough so that it starts to think the nebula is formed of stars.
This is a good star mask, notice the strong red on bright stars and the galaxy details free from protection.
Too weak star mask, the stars won't be protected from deringing. Only the "super bright" stars are semi-protected.
Over-protected deringing mask. Notice how the background and the galaxy details are considered stars!.
2) High SNR mask creation Create a mask for the high SNR areas. You'll have to tweak it until you see that it only deconvolves the zones you want
A good mask example. Notice the small stars are included, the background is excluded and there is a smooth transition between the highest SNR zones and the lowest. This mask is intentionally inverted to show better the protected zones, you must not invert it, otherwise you'll be only deconvolving the noisy background!
3) DPSF Analysis Analyze the stars profile with DPSF. Take into account that this is not going to work in all situations. It tends to generate a standard deviation set too high.
A DPSF example. These values will be used for the final deconvolution process. Take note of them.
4) Go ahead!Protect the original image with the mask. Activate the local deringing with the star mask, and disable the global dark deringing. Start with 50 iterations.
Example of starting deconvolution. Notice the increase in resolution.
5) StdDev tweaking Lower and Raise the standard deviation, inspecting it visually. The perfect value stands in the middle of starting to get only small scale noise (too low) and starting to get high structures, detail loss and black artifacts around stars (too high).
Example of StdDev modification, again, check for halos and detail improvement.
6) Iterations, moise reduction Tweak, again visually, the number of iterations and the noise reduction. Do not try to get a lot of iterations and a lot of noise reduction. There is a point where you cannot get more from the data you already have.
7) Deringing global tweaking If there are slight black halos around stars, try to increase a bit the global dark deringing. But don't lose detail in the nebula!.
Example of too much global dark. Notice the bright artifacts.
Q: I have some tilt in my optical train, can deconvolution solve my problem?.
A: Yes, partially. You can model the shape of a star (rotation and elongation). You'll have round stars at the end. Use the DPSF tool to measure them.
Q: I have some guiding issues, can deconvolution solve my problem?.
A: I'm sorry but no. Although you can improve a little bit with some motion blur deconvolution.
Q: My deconvolution adds a lot of noise!.
A: First, did you protected it with a mask so that only the high SNR zones are deconvolved?. If yes, there are two possible scenarios:
1) Per pixel noise: you have to increase the noise reduction in the first wavelets scale. Also check the rejection algorithm while stacking.
2) Rugosity-like: Either the mask is not working well, you have too much deconvolution steps or you have to increase the second scale wavelet noise reduction.
Q: My stars have black halos, how do I solve that?.
A: Check this steps:
1) Did you use a star map for the local deringing map?
2) Did you use a mask to protect the high SNR areas?
3) Try to increase a bit the global dark in deringing. This is going to work for sure, but you have to do it and avoid the generation of artifacts in non-stars objects.
4) Your standard deviation is set too high, try to decrease it. DPSF generates oversized PSF sometimes.
Q: Richardson-Lucy or Van Cittert?
A: Richardson-Lucy works best for deep sky images, Van Cittert reaches the same deconvolution result with less steps, but it tends to generate more low scale noise, so you'll have to fight with the first wavelets noise reduction a lot more.
Q: But I want more sharpness!.
A: You can try Restoration Filter and ATWT on the zones with the best SNR once you apply a non linear histogram transform.
Q: What is my standard deviation?. Is there a magical number, or way to calculate it?.
A: Yes and no, the best is to find it inspecting visually. If your seeing is almost the same, this value is fixed. If there are some small deviations beetween days, and you stack different data from a lot of days, it also tends to be fixed, because mean value of the seeing for one site, in long periods of time tends to be the same (if you collect a lot of hours like me).
Q: You are doing it wrong!. Why don't you use the DPSF?.
A: Well, I have experimented with it a lot, and I really prefer my version, it has the most round stars, and looks more natural to me. The DPSF version has black halos, and correcting them with global dark creates bright artifacts. What would you choose?.
DPSF vs. my way of doing it