ngc7000, deconvolution help

mmnb

Well-known member
First time I think I have an image worth deconvoluting. Decent amount of exposure and I think the stars look good and round (I was having some problems with sensor tilt earlier). This is the linear image after PCC and automatic background extraction (there is some annoying green in the corner, not sure where that is coming from, but wanted to crop at last step):

PSF:

Star mask (made with MaskGen):

I remember a talk from Adam Block a while ago (AIC in the before times) where he tried to demystify deconvolution a bit and he talked about how you should be really looking to see if the stars are getting smaller. The only way I get that to happen on my image is if I turn off deringing using just the star mask to protect the the star centers (global dark=0.008., default wavelet regularization settings).

Can anyone help provide some direction on how tell if the deconvolution is doing what it is supposed to?
 
in this image i think because the star sizes vary so much from channel to channel it might be best to split it into the 3 color planes and compute PSFs for each channel separately, and deconvolve each channel separately. i think each channel is going to need different deringing parameters as well.

overall the SNR doesn't seem so high so you probably clone each channel's image, stretch them and use them to protect the lowest SNR areas of the images during deconvolution.

the only problem with this method is of course if you don't end up with similarly-sized stars in each deconvolved image, the reassembled RGB image is going to have colored star fringes. but they are already sort of there in the initial image (purple fringing)

perhaps with this OTA/camera you need a UV block filter...

rob
 
Yeah the stars are a lot bigger in blue....I think I have the focus offset right. I see some people using UV/IR cut filters for the QHY 16200A.

I will split and try channel specific PSFs and masks...can you provide any other guidance as to what cues to look for in the image indicating "good" deconvolution?
 
i guess the below are actually signs of bad deconvolution...

since you probably need to turn on deringing, the main thing is to make sure that you don't see the message about divergence in the log as the iterations are going. that means the deringing global parameter (or local parameter) is set too high. i usually start with 0.1 on global and if there are lots of divergence warnings and the image is completely destroyed, i back it down to like 0.05 and keep going that way until the image looks OK.

one thing that tends to happen is that the energy in the star is spread out due to the convolution that the atmosphere performs on the image and so when you deconvolve all that energy is pushed back toward the center of the star. this can cause the center of the star to be blown out, but a carefully constructed star mask can prevent that. of course now you have the problem that you want to mask the low SNR background *and* the cores of the stars, so you have to use pixelmath to make a mask that protects both. in theory also you could use the dynamic range expansion controls way at the bottom to give more "headroom" at the high end but i've never really tried doing that.

finally sometimes deconvolution can give your images "the wormies" - hard to describe but you end up with all kinds of little twisty structures that weren't there before. you can kind of blink the before/after with the undo/redo controls to see if those are showing up.

so i guess good deconvolution is when you have the absence of the above stuff and the stars and structures look visibly sharper.

rob
 
I extracted the red channel and made a specific PSF and used a star mask as a control. It feels like the opposite is happening, the stars look bigger after deconv. If anyone can show me what, say, just the R channel *should* look like post deconv that would be much appreciated.
 
if the stars look bigger then probably the deringing control is too high. try turning off all deringing and see what you get (you'll probably have black rings in some of the larger stars, but you should see the smaller stars shrinking.

rob
 
Yes, I did that at the beginning (but did it again to be sure). Turning off deringing entirely does make he stars ever so smaller...the dark rings, when you can see them, seem to be tolerable. Does one just take the result with deringing off when you see this?
 
i've generally been able to get rid of dark rings with fine adjustments to the global dark setting - what value are you using?

deconvolution in general is a black art since it's not a "well-posed" problem. there are a million different "true" images that might correspond to your convolved image, and so there's no right or wrong settings when trying to deconvolve, other than settings that result in images you don't like. sometimes it's just really hard to get an acceptable result.

rob
 
The stars grow subtly even at global dark = 0.008.
Turning deringing off seems to give the cue (smaller stars) that makes sense. I can see a fairly clear improvement in sharpness in the nebulosity.
The ringing that I can see up close isn't very visible at the intended image scale and is perhaps ok:
"Also take into account that a very small amount of visible ringing can be beneficial rather than harmful, since it increases acutance, which in turn favors the visual perception of detail."

I'll stick with that and post finals and see if anyone is terribly revolted by what I end up with.
I'm also not terribly sure if the separate channels matter, I am making the PSF from the extracted L channel...so, should be fine enough?
 
well what i was seeing when i deconvolved your image was good deringing in 2 channels and bad in one (G if i remember right), so it seemed like a good idea to separate the channels in order to apply different deringing parameters. you can probably get away with the same PSF for all 3 channels but it might make sense to try with customized PSFs per-channel.

there's a script by HVB that automates the PSF extraction process. just make sure the number of stars it finds is sane - probably 50-100 is more than enough to get an accurate PSF.


rob
 
Back
Top