I reprocessed my M101 data (20h RGB + 2.5h Halpha) from march to test some new processing techniques. Since the final result has become a completely new Image, it deserves a new upload and it's own description. The main focus of this reprocessing is the sharpening as well as the H-Alpha integration. Let's start of with the sharpening:

1. At first, I created a revision with only sharpening based on stretched 16bit data in Photoshop.
2. The next revision contains a previous sharpening in APP on linear 32bit data, with consecutive sharpening in Photoshop 16bit data after stretching.
3. the final revision shows the BlurXTerminator deconvolution in PixInsight on 32bit linear data, with an additional slight (selective) sharpening of the spiral arm structures in Photoshop

One can clearly see the tremendous improvement of details over those methods. The following gif shows an even better visual impression of of those improvements. It shows the comparison of ordinary Photoshop sharpening vs @Russell Croman's BlurXTerminator:


I was wondering if those incredible results can only be achieved with such a long exposure time (20h), or if it also works with less... Therefore, I created four different stacks of the same data, with different accumulated exposure time: 2h, 6h, 12h, and 18h.



One can see a crop of M101's core with different exposures and activated/deactivated BlurXTerminator. Those images are the raw stacks with the same stretch settings. Astonishingly to me, there is barely an improvement of details between a 2h and an 18h exposure! The spiral arm structure looks equally amazing, even with less integration time. Only the darker background areas show an improvement in SNR with better exposure time (therefore allowing to "catch" the faint outer parts of a DSO).

My conclusion is that BlurXTerminator is a complete game changer for my processing. I am not very acquainted with PixInsight deconvolution, therefore I can not evaluate if such results could have been easily done with other plugins.... But I will definitely work with this new AI from now on :-)

The second processing improvement is the H-alpha integration. I usually do it in Photoshop, where I "screen" the colorized Ha channel on the RGB Layer. But the Ha-stack is very noisy and usually spreads a "red glimmer" over the whole core area. This time I used the continuum subtraction technique, where I subtract RED - HALPHA. This leads to a way better representation of the hydrogen areas since other luminosity that also emits in the red spectrum has been subtracted.



EDIT: for those who wonder "does BlurX add fragments to the image that aren't even in the data?" Short answer: No.
In order to do this, BlurX must plate solve every image and replace fragments of our image with parts of a gigaaaantic database. Which is simply impossible for a tool with a size of several MB and no demand for an internet connection.
The secret is a carefully trained machine learning model based on quality ground truth data. We have a tremendous amount of observations in our raw stacks, that we can't even comprehend. From my understanding of general ML approaches: The AI decides for small areas (e.g. 6x6 pixel) if it's signal or noise and evaluates the pixel correspondingly for each patch. It can do so because it's trained on a large set of data with superior accuracy (maybe such as Hubble or JWST). It works the same way how AI can analyze an MRT scan better than senior radiologists - because the AI model is able to draw statistical conclusions that we simply can not comprehend. So AI based convolution isn't cheating - but it's a revolutionary processing method, and hopefully not the last to come.

Happy New Year 2023