Removing stars with StarNet++

To cleanly fold Ha data into our LRGB images, we want to remove all stars (otherwise the star colors will be distorted). In the past I tried various options:

• Creating a star mask and subtracting that from the image. Which leaves black holes in the image. It's not too bad as these black holes are where stars are. But if the stars in the star mask are not the same size then the real stars, some artifacts are left.
• I tried Straton - but could never get a clean image out of it.

So, I was exited when I read about StarNet++ - an ML-based algorithm that removes stars. And it even comes with the latest version of Pixinsight:

First time, I wasn't sure what these "weights" files are... Turns out, that these are files Starnet files that DON"T ship with Pixinsight. Weird.

If you go to the SourceForge project of StarNet you download the whole standalone software, unpack it and have the files in there:

    mono_starnet_weights.pb

    rgb_starnet_weights.pb

You copy them into the Pixinsight library folder (or a subfolder), click on "Preferences" and select them:

Now, one challenge is that this process only works on non-linear images - but we want to fold the Ha data into the RGB data in the linear stage.

I found this video where somebody explained how to create a "reversible stretch". We use the ScreenTransferFunction:

And click on "Edit Instance Source Code" (the little square icon at the bottom):

The middle value in the first row (here: 0.0051) is the midtone stretch factor. We use that to create a PixelMath stretch function: mtf(0.00051,$T)

When we apply this to our image instead of the ScreenTransferFunction stretch we get a stretched linear image that looks like this:

Not the best contrast, but enough to let StarNet do its work. If we now apply the StarNet process to this image, we get this (this can take quite a while!):

It's a little bit hard to tell, but the stars are gone. We change the expression in the PixelMath window to: mtf(1-0.00051,$T) which reverses the stretch. If we now stretch the image again with ScreenTransferFunction we see that the stars are gone:

The larger stars leave some shadow behind, but that won't matter too much. We can use this image to fold the Ha data into the RGB data. I use the NRGBCombination script (under Scripts->Utilities) for this.