Starnet++ training for reflectors (Let's collaborate!)

[Leeps]

So I just started to try out Starnet++, and it's a super cool way of using a convolutional neural network to remove stars. The issue, as stated by the creator, is that it's trained using a dataset from a refractor scope, and so it doesn't work that well on reflectors, or any other scope type that has diffraction lines or any other focus aberration. 

So to train the AI, we need as many images as we can get our hands on, both preprocessed, and after having the stars removed, but shot with a reflector scope. I wanted to see if anyone is up for collaborating on it, so that we can get a lot of data together, and crunch it. The more we get the better the removal process will be. 

If you want to help, here's what we need:

1. Process your photo normally (histogram stretch etc.)
2. Export it as an 8 bit TIFF
3. Remove the stars using any good method (I believe there's a few about, but this one seems to work)
4. Upload an 8 bit TIFF of the original processed image and the star removed image, clearly labelled
5. Drop the link in this thread!

If we can get a few dosen images, we can train this thing and see how it does. I'm using a 200p, and probably have 10 or so images I can use. We're only likely to need a couple dosen. After that, if anyone else has any other systems that Starnet doesn't work well on, we can try those. Ideally for this first batch, Sky-watcher P or PDS models (200p, 130p, 150p etc) should give us enough data if we collaborate. 

Up for it?

[Leeps]

To be clear what the issue is, this is how Starnet++ is handling my images at the moment: 

 

It doesn't know how to handle the diffraction spikes or the coma error. I'm hoping to help fix this with the retrain! If it can be shown coma stars and diffraction lines, it can learn to remove them. 

[Leeps]

As an update to this, I managed to get the python versions correct, and have the neural net training on my own data. I have about 5 images that I de-starred and set as the training data, and have run 30 "epochs" so far. Aiming for around 100 before I retry the crescent nebula again. 

Anyone that can help, even with unedited, good examples of diffraction spikes that I can edit would be great!

[Allinthehead]

 

[johneta]

Hey that's a great idea good luck.... Can I just question it a bit?

If you can do good star removal by the method in the link you mention, then why use Starnet.
I only mention this as I always struggle with star reduction and getting artifacts. Ive tried Starnet (I also have a 200mm newt) and have had limited success.

If you train it with starless images with artifacts, aren't you just training it to produce starless images with artifacts, which are then hard to deal with?
I know its a bit of a circular argument I'm expressing but I guess I don't quite get it.
Ignore this if its too off topic or hijacking the thread.

[Leeps]

No hijack at all, glad to generate the conversation. The idea is that you have to unfortunately manually edit the photos to remove the artefacts. The nature of the beast is that it's doing a removal and replacement of things it has seen before, and it'll never gove you a "true" image from a data sense. So I've been editing files to remove stars mainly using photoshop's content aware fill and healing brush. This gets you to a clean looking starless, but it's a bit of work! 

The idea then is to run the ai which does 90% of the work for you, and may be able to handle scenarios that are hard to manually edit better. If there are artefacts, you'll still have to briefly manually edit them. 

I'm learning as I go, so far I've managed to make it much worse 😁 but the more training data I can get, the better the chances are. 

[Leeps]

On 05/07/2020 at 14:44, Allinthehead said:

 

*Pics*

 

This is awesome. Do you have usable 16 bit  TIFs of the above? 

[johneta]

I understand what your saying there. My issue is with areas of nebula that are buried in the milky way or close by, which are an absolute mass of stars. It is extremely difficult to remove the stars in that kind of image by hand, and to therefore feed it to the Starnet Algorithm. I don't have an issue with images such as the excellent Horsehead one above, where there are a medium amount of stars which are easy enough to deal with by normal methods.
Also some of the  Starnet 'starless' images I have used are of quite good quality, but I still have difficulty getting a natural looking image in the end trying to blend stars back in.
I guess I'm just frustrated with Starnets wow factor, and my inability to get anything usable out of it.

Again, good luck. It seems like a worthy objective your shooting for.

[Leeps]

 

These are the files I'm currently training with. I'm not getting great results, I think I need more nebulae that is background... I'm pretty new to shooting this stuff, and so that's why I was hoping to draw on other's images if they'd share I don't have many shots of H-alpha nebs. I think a good shot of the rosette would be good, or similar. 

 

Edited July 8, 2020 by Leeps

[Allinthehead]

On 08/07/2020 at 17:53, Leeps said:

This is awesome. Do you have usable 16 bit  TIFs of the above? 

Here you go.

https://www.dropbox.com/s/lropbcjtpycp06f/starless.tiff?dl=0

https://www.dropbox.com/s/6h8fok57715l2hi/horseheadfinal.tif?dl=0

[Leeps]

On 11/07/2020 at 13:06, Allinthehead said:

Here you go.

https://www.dropbox.com/s/lropbcjtpycp06f/starless.tiff?dl=0

https://www.dropbox.com/s/6h8fok57715l2hi/horseheadfinal.tif?dl=0

Thanks! I'll incorporate it into the training. 

 

After training for a while on the images above, I'm finding fairly poor results so far. If you're on this thread and you have usable images, I'd love and appreciate the help! We can make this thing better for everyone with a reflector

[rnobleeddy]

Just a quick bump to see if anyone made any progress with this?

I'm familiar enough with the underlying technology, so the barrier is just going to be data. I'll have a go at a couple, but I haven't found any tool that does a good job of picking out stars + diffraction spikes, so it's going to be slow going!

[Davey-T]

Is it not easier to just remove the diffraction spikes before removing the stars ? I think APP can do this.

Dave

[1stSolo]

AstroPixelProcessor removes diffraction spikes by default, unless you chose to protect them. After that standard StarNet++ should take care of star removal.
However I'm also interested in training starnet to do its job better. I almost always have artefacts left after it that I clonestamp in PI and it's not always looking great. I wish there was an API for community to update .pb files with new training data to benefit everyone.

How far along are you with training and how are you planning to update the .pb files?

[Chefgage]

On 16/01/2021 at 11:17, rnobleeddy said:

Just a quick bump to see if anyone made any progress with this?

I'm familiar enough with the underlying technology, so the barrier is just going to be data. I'll have a go at a couple, but I haven't found any tool that does a good job of picking out stars + diffraction spikes, so it's going to be slow going!

Tried it a few times with not very good results. Seems to badly affect the nebula/galaxy when removing the stars. Also exactly how does this star removal work? By this I mean when I have got it to work if I then do another levels/curves stretch (of the starless image) I can bring the stars back out?? So does starnett++ just cover up the stars or is it just not working for me?

Edited May 21, 2021 by Chefgage

[michael.h.f.wilkinson]

I actually have a student working on a very different, very fast approach to star removal. It would be interesting to see how that fares on reflector data. Ideally we should have a prototype running by the end of next month, but to date he only has data from my 80mm refractor and 6" Schmidt Newton available, so no diffraction spikes

[dweller25]

Why do you want to remove the stars?
In my opinion the images with the stars look far better 🙂

[Chefgage]

28 minutes ago, dweller25 said:

Why do you want to remove the stars?
In my opinion the images with the stars look far better 🙂

It's so that you can process the image/nebula/galaxy without the stars being effected so like using a star mask.

[AstroMuni]

Sorry, Not sure if this is right thread for it but I am struggling to get Starnet++ to run on my old Linux laptop. I get error 4393 illegal instruction when I run it 😞 I am no expert on Linux or Python and a google search left me no wiser... Anyone else who can shed some light on this please?

I am on Linux Mint 20.2 Uma.

Edited July 15, 2021 by AstroMuni