›

Starnet and Narrow Band and scars

Linwood
in PixInsight
Is there any real fix to starnet's poor performance on narrow band images? 

Here is a 1:1 view of an otherwise unprocessed SHO combination from PacMan, left is before, right is after starnet.  Some stars are removed and leave pock-marks behind, some are just ignored, especially smaller ones.  I am more concerned by the pock-marks or removal scars really than the small unprocessed ones.

I have had poor luck getting selective enough with the alternative of wavelets for removals, though at least there is some tuning to do there.

Is there any tuning for starnet?  Any replacement files, for example, with different training? 

image

Comments

  • Gordon Aspin
    You could try the technique here https://www.chaoticnebula.com/pixinsight-star-removal-with-starnet/

    I haven't used this approach myself but it may help with bigger stars which seem to be your main concern.
  • BlockHeadBlockHead
    A couple of things:

    1. If you intend to put the stars back after adjusting the starless image- the "pock-marks" will not be a big issue. They are really only an issue when you are just displaying a completely starless image.

    2. No, I do not believe StarNet has had any updates to its training. So... you will always get these kinds of results- only new/additional training would make a significant difference.

    3. Given that... you might want to investigate Russ Croman's StarXTerminator. He is actively still training and he goal is to always do better than StarNet in general. 

    4. My typical use of a starless image is to make very faint things brighter. This isn't the only use...but when it is, a very aggressive noise reduction is fine since it isn't detail that is super critical...it is just seeing the stuff.

    -the Blockhead 
  • Linwood
    Sorry for delay, I thought I would get notified of responses, but didn't.... 

    The latest reason for this is to replace the NB stars with RGB stars, but substantially reduced RGB stars as well.  Where the RGB stars are large that works, but if I tune those down (which is fairly straightfoward) often the pock mark remains behind it.  I've watched the video in using halos from the starless to back fill around smaller stars, but in this case the starless image has the pock mark.  This was a handy example, but I have had others (like a LOT of data on Bubble) that the scar is even more prevalent. 

    StarXTerminator -- off to find it and give it a try.  Thanks.

    I also see you have an LRGB -> NB star section in horizons.  Have hesitated subscribing as that seems about the only NB content there, and I am doing more and more NB just because of light pollution here.  Is more NB stuff under the covers there?  Or do you plan more? 

    I got about 65 hours on the bubble nebula, then spent two or three times that banging my head against the desk (well, not literally, but it felt like it) trying to get good results.  Even just choosing a good NB pallet is tough, much less making it all come out not looking like mud.   The starnet issue is because I sometimes get pock marked mud.  :(


  • BlockHeadBlockHead
    I plan more NB.
    Just can't produce videos fast enough...
    The SQUID nebula is in process.


    Regarding the pock marks. You can definitely use my technique that I employ in the MANUAL version of my star deemphasis. Have Starnet (or StarXTerminator) output the stars only image as well as the star image.
    Use the stars only image as a mask on the starless image. Then use MMT with a large wavelet layer to fill the star regions. Finally add back noise that matches the "grain" (texture, pixelation) of the starless background.

    Your scars will blend in a little better if you do this gently.


    -the Blockhead
  • Linwood
    Thank you I will try.
  • Andrea GironesAndrea Girones
    thanks for this Adam I am also new to narrowband and am going crazy with my weird stars and artifacts!  Can't wait to see the squid nebula. I imaged in in June and  had a  great time ( no not really) processing it.

Sign In or Register to comment.