AI for Astrophotography (2022 AIC Presentation)

By gracious permission from the board of the Advanced Imaging Conference, this is a presentation I did in 2022 introducing convolutional neural networks for astronomical image processing. The main goal is to demystify AI and neural networks and show you how they work.
Check out more great AIC content at www.advancedim.... All content online can be accessed by creating a free account.
This video is copyright © 2022 by the Advanced Imaging Conference. It may not be reproduced in whole or in part without prior permission.

[@cosgrovescosmos]

Russ - Thank you so much for posting the video of your talk! This is an outstanding overview for those that want to better understand the AI Network technology that you are using to create the tools you produce. I can say that each of your tools has fundamentally changed the way I process images! This video was posted after I had already written my recent article on BlurXTerminator - which is a shame as this would have been such a wonderful resource for me as I was pulling that together. In that article, I tried to give a high-level overview of the technology to help people understand it and to dismiss some misunderstands of it. This does a beautiful job of doing just that. The good news was that I was not too far off the mark, and I will be updating that article with a link directly to this excellent presentation. Thanks again, and keep up the great work!

[@rrcroman]

Thanks much, Patrick. I very much enjoyed your in-depth article. Thanks for taking the time to do it!

[@cosgrovescosmos]

@@rrcroman Your welcome Russ - I will do the same for your follow-up projects. You may not be able to keep Chocolate in your pantry anymore, but you changing how image processing is done for Astrophotography, which seems to be a worthy trade-off! On the other hand, I now feel compelled to reprocess all of my old images as I know I can make them so much better now! All the best, Pat.

[@michael.a.covington]

38:00 Extremely important point -- the neural network is NOT a model of conscious thought or even explicit knowledge representation. It's patterns of numbers. I wish more people realized this.

[@jonathanpearceff]

Great presentation explaining the concept of how the XTerminators work. I really liked the short video of the network learning on an image.

[@scottbadger4107]

What about AI for autoguiding? Seems like a great application for it. RMSXterminator.....

[@duartefaria7134]

I just wanted blurXterminator for photoshop. I do not want to spend 300 more vat for pixinsight. Your software is awesome, i realy love the results, i already have noiseXterminator and gradientXterminator, only missing that one. Please, free us. Photoshop is a software that i use in my professional life, i will never ditch it. I just want to spend less. And i think i talk for anybody with the same mind-set. If you do not make it for money, make it for the community.

[@scottbadger4107]

Great presentation! Sorry if I missed it, but is the down-sampling and consequent up-sampling done simply to manage the computational load, or is there some other reason?

[@rrcroman]

This is done to increase the "receptive field" of the network. Each layer consists of relatively small convolution kernels, often just 3x3 pixels, so it can only "see" features within that kernel size. Downsampling followed by convolution gives each successive layer 2x the receptive field. By the time we get to the inner layers of the network , they can "see" relationships between features that are far apart in the image. This information is then fed to the back-end of the network through successive up-sampling and convolution, so the network can "put it all together" and make decisions about what to do that includes information across many scales.

[@scottbadger4107]

@@rrcroman Makes sense - and interesting - thanks Russ! And please keep Xterminating!!.....

[@MikeViker]

Clever presentation!

[@bullshitvendor]

Would applying BX to your subframes be of any use or just the final stack?

[@scottbadger4107]

I actually tried that with 40 frames of M51, but BX on the integration was better. Probably because of a higher SNR.

[@keithhanssen7413]

👍🏻

[@rodjos5463]

👍

[@bullshitvendor]

oh yeah, I forgot, a ComaXterminator would be awesome! I could get rid of all those extra glass paths and a 1000 grams of weight hanging off of my focuser