Hi, everyone! I am finally releasing an updated version of StarNet - StarNet v2.
The link to download is here. Make sure you are downloading the right version.
I know that this release is long overdue, but I will touch the reasons behind the delay in the last section of this post.
This version took countless hours of work tp prepare more training data, with the push towards star dense images. It was also trained using hardware better than my 10 years old laptop (now deceased...). It should show better performance across the board, including wide field images of the Milky Way (well, also wide field images with no Milky Way in them, if anyone have ever seen those, lol). Hopefully, this will enable StarNet usage for landscape astrophotography. This is ultimately for community to decide, but it works pretty well on images from my rigs. This version should also work well for images at any stage of processing, not only after initial stretch, which makes it more useful.
Good news for people who don't have PixInsight - this release includes simple user interface, so you won't need to mess with command line and other scary stuff (hooray!). The negative side of this, though, is that GUI is only available for Windows. I am not able to compile GUI for MacOS, and most likely never will. I know there are ways to run Windows apps on Macs, but don't ask me how, I have no idea. Hopefully the community will be able to help here.
This release also includes new version of the PI module. PI team will probably include this into PI distribution at some point, but right now they are busy with other things. This release will install StarNet2 separately and will not replace the original StarNet. The installation procedure is pretty much the same as before and is described in the README.txt file.
A few notes on updates and changes:
- It was a huge mistake to include a large number of stride values available, and many people wasted a lot of time trying them all out. The new version has only two (256/128) and I highly recommend leaving the default one (256) on!
- Smooth tile transitions will now reduce the number of artefacts.
- There is support of linear images (please read on). The neural network itself does not work with linear data, so this feature (if enabled) will use STF parameters to stretch the image first, remove the stars, and then de-stretch the image back. Just make sure to get STF correct before running StarNet. I suspect this will lead to disproportionally large amount of discussion around it with purists arguing how it is not really a linear image anymore after you stretch and de-stretch it back and so on. Still decided to include this, despite people recommending me not to. Please don't make me regret this, don't like it - don't use it.
- There is also 2x up sampling option, which will be explained below.
Known problems:
- Standalone version might loose your color profile, so the colors might appear different. If this happens, all you need is to just open the image in the software of choice and restore the profile.
- The most annoying problem is that sometimes this version will leave visible patterns around small stars on some images. This tends to affect images with high SNR and tight bright stars. To combat this I recommend up sampling your image (2x). This removes the problem at the cost of higher processing time. There is no option to do this internally in the standalone version, so please upsample your image in PS or other software you use. This problem really bugs me out, but so far I could not get rid of it.
I want to say huge thanks to Rob Pfile, Alberto Ibañez and Eric Coles for their continuing support! This release would be impossible without your help!
That's about it for the main part of the post I guess. I hope there will be no problems with PI installation and usage, but the new GUI for Windows might potentially be problematic. I have never built GUI before so I don't know what to expect, but it might not work for some people. Any help will be appreciated.
Now a bit on the delay reasons and new license term conditions. Apart from the fact that I started a new position recently and have little time left for my hobbies, including AP, there is another reason behind the delay. This updated version has been sitting on my laptop for a while, and I was not sure how to go about releasing it. I don't know how to describe this in simple terms, but the problem with releasing it is that I have no tools to protect my intellectual property. Conventional software is pretty safe in this regard because once compiled, the original source code is no longer easy to recover. With neural networks based software the situation is different. I spent countless hours manually photoshoping images full of stars (imaging removing stars from wide field image of MW) to get training data. One reason behind this was the fact that there was simply no software out there that could help me (exactly why StarNet was created). However, once such software exists (StarNet), there is nothing that can stop someone else from using StarNet to create training data and train seemingly independent neural network. There are some published research in this area that provides ways to protect intellectual property and I tried to incorporate some of those into StarNet, but I still have no guaranties whatsoever that this will help.
Long story short - this release comes with updated license terms that specifically limits StarNet usage to astrophotography processing only. In no form shall StarNet or any of its components and outputs be used to create any derivative works, including but not limited to, training of any kind of neural networks for any king of image processing purposes. This includes not only star removal, but any other kinds of image transformations. StarNet shall not be used to aid development of other software. This limitation includes some non-trivially harmful actions like sharing (PMing) original images on which StarNet works well (and other software does not) with interested parties. Overall, just remember that if something is given to you for free doesn't mean it has no value and you can do whatever you want with it.
The open source part of the story is also over, so the code on github will also not be updated, unfortunately. I will also not reply to any inquiries about the details of StarNet's development/training etc. due to reasons I won't disclose now. Any information about StarNet's NN, training/test datasets etc. is now proprietary.
Anyway, I really hate writing about these formalities, but I feel that I have to. Above all - just have fun and create cool images!