StarNet
StarNet started its life in 2019 as a script known as StarNet++ written for PixInsight by Nikita Misiura. It was featured in PixInsight’s Script menu, which incorporates many free third-party plug-ins. Whereas the unique model did a commendable job of eradicating stars, it took a quite very long time to run and left some star artifacts behind.
Regardless of these minor shortcomings, PixInsight’s developer, Pleiades Astrophoto, embraced the device. The corporate transformed the plug-in from JavaScript to C++, the identical language as PixInsight’s most important processes, and integrated it as a local device in the primary Course of menu, the place it’s now known as StarNet.
The first purpose for the less-than-perfect efficiency from StarNet++’s unique algorithm was a scarcity of coaching of the neural community. The important thing to machine studying is coaching algorithms on huge, diverse portions of knowledge. The looks of stars can range tremendously in brightness and form, together with halos and diffraction spikes. Whereas the human mind is aware of that every one these variants are stars, a neural community should be taught to acknowledge them.
Misiura’s unique algorithm was skilled utilizing solely knowledge taken along with his modest imaging system. And when he first launched StarNet++, he was finishing postdoctoral work in knowledge science, leaving little time for enhancing this system. In early 2022, nevertheless, Misiura launched StarNet2, a vastly improved model of the unique course of. (On the time of this writing, it has not but changed StarNet 1.0 within the PixInsight common launch, however needs to be included in a future revision.)
model that may be run from a command line or graphical consumer interface on Home windows, MacOS, and Linux.
Whereas it’s advisable that stars be eliminated early within the workflow, I often make some fundamental corrections to the information earlier than making use of StarNet2.
A consultant workflow may embrace the elimination of sunshine gradients, shade correction, linear noise discount, delinearization (nonlinear stretching), and a contact extra noise discount after the stretch. In PixInsight, instruments to perform these duties is likely to be DynamicBackground Extraction, ColorCalibration, MultiscaleLinearTransform, MaskedStretch, and TGVDenoise, respectively.
When processing knowledge from a shade digicam (one-shot shade, DSLR, or mirrorless), it’s superb to first apply a contact of shade saturation with CurvesTransformation after which run StarNet2. If working with monochromatic knowledge that features luminance, it is smart to attend till the luminance knowledge have been mixed with the colour picture utilizing LRGBCombination — on this manner, star elimination want solely be executed as soon as.
StarNet2’s interface is straightforward however has just a few choices. Depart the Stride setting on the default of 256. In case you intend to interchange the celebrities after processing the starless picture, select “Create starmask.” It is a little bit of a misnomer — the picture produced isn’t a masks in any respect, however quite the celebrities themselves (proven in step 2 beneath).
Misiura states that some “tight and shiny” stars can produce artifacts, and the “2x upsample” choice, which upsamples the picture earlier than star elimination, is designed to alleviate them. In my checks, I felt this was pointless.
I don’t advocate utilizing the final choice, “Linear knowledge.” This internally delinearizes a linear picture, then removes the celebrities earlier than unstretching the picture again once more. Greatest follow is to make use of a picture that has already been delinearized with MaskedStretch, as described above, or with HistogramTransformation.
After working the plug-in and isolating the nebulosity, you may therapeutic massage high-quality particulars through PixInsight’s full array of instruments with no threat of blowing out the celebrities.
Step 3 beneath demonstrates how properly StarNet2 performs on a picture stuffed with 1000’s of tiny stars. The end result has additionally been enhanced with the processes seen within the Historical past Explorer: HistogramTransformation and ExponentialTransformation to brighten the picture, CurvesTransformation for enhancing shade saturation, LocalHistogramEqualization for enhancing distinction, MultiscaleLinearTransform for sharpening, and HDRMultiscaleTransform for bringing out the element within the cores of the nebula by compressing their brightness. In fact, there’s hardly ever a technique of skinning the proverbial cat, and an identical end result might be achieved with a special combine of accessible instruments.
As soon as you’re glad with the event of the nebula, it’s time to change the celebrities. You may as well make delicate changes to the picture that incorporates the celebrities — like tweaking shade saturation — earlier than reuniting it with the primary picture. Utilizing the settings proven in Step 4 beneath, apply PixelMath to the starless picture for those who want to put the celebrities again in. You would additionally attempt lowering the star contribution by multiplying the star-mask picture by a quantity lower than 1 (e.g., starless + 0.75*star_mask).
