Wow, absolutely top notch presentation, Russ! You took a complex set of concepts and did a beautiful job at bringing the audience to that "ah ha" moment where I can clearly see the power of combining Drizzle with Deconvolution, considering the Point Spread Function. A wonderful journey ✨️ and I will definitely apply to my wide field / undersampled / 5-7" per pixel image scale images to improve contrast and resolution 👍 Thank you and keep looking up and enjoying the journey... Cheers!
Fabulous presentation by Russell… i remember signals and systems undergraduate classes not-fondly… from 42 years ago. Russell presented the information in a way that I might have enjoyed those classes back in the early 80’s if I had such a talented instructor who applied the theory to practical uses. Unfortunately I had the graduate student teacher who was terrible and ruined the experience. Luckily semiconductor physics caught me and I had something I could use in my semiconductor focused career! Russell is really upping the game of so many astrophotographers. Thanks Russell!!!
Fantastic presentation. I dont currently use BXT, but will definitely be looking into it. I've recently decided to make the effort to truly understand the functions of the various processing tools, and this presentation has gone a long way to helping me understand deconvolution. Thanks Russell.
That was a terrific presentation leading to several "light bulb" moments. As a relative newcomer to astrophotography and Pixinsight it was invaluable in aiding my understanding of what happens twixt star and pixel
You did a great job presenting deconvolution and hopefully put to bed some of the criticism that "BXT" creates data. You did an even better job in building and training BXT! I'm using an 11" ~F/1.9 scope at about 1.5 arcsec/sample.. The optics introduce some off-axis aberrations including chromatic. BXT does a great job at correcting the optical abbs and reducing seeing, most impressive, but nothing compared to what comes next. I reprocessed a couple sessions using 2x drizzle and a .65 spot factor(underwhelming) but I was totally blown away by the BXT output! The low contrast non-stellar improvepment was even more dramatic than the stellar. Detail grew magically (but faithfully) from the mist.
I learned a lot from this presentation, especially on “classical” signal processing. Are there any presentations going into more detail on the neural network / ML side of BXT?
Beautiful and all and thank you. I bought BlurXt, but I am not impressed by the sudden price increase, from $60 to $100, and from 40 to 90 for the second app. I understand we're living in a cosmic inflationary period but 67% increase seems excessive.
excited about what he has achieved. I have a question Russell ... is the system good enough to perform photometry? Is the data lost in the background of the image restored correctly? Regard
From the BlurXTerminator documentation: Great care has been taken in the design and development of the neural network architecture, and especially in the training methods employed, to try to ensure that BlurXTerminator produces results that are faithful to reality and not "inventive." It is not and never will be perfect in this sense for all images produced by all instruments. Star centroids may not be perfectly maintained. Total flux may not be perfectly conserved. Exact shift-invariance (getting the same result regardless of whether an image is translated by some number of pixels) is not guaranteed. BlurXTerminator's "best guess" at what detailed structure underlies the original blurred image may not be perfectly faithful to reality in all cases. A consequence of this is that BlurXTerminator should never be used for quantitative or scientific applications. Neural networks and other forms of "machine learning" are indeed finding applications in scientific astronomy, but most often in the form of tools fine-tuned for particular instruments, not general-purpose solutions like BlurXTerminator. Their results are still looked at with a bit of a jaundiced eye, and surprising or potentially important results are always cross-checked using traditional, more transparent methods.
High spatial frequency is detail (short wavelengths), and very high spatial frequency is generally noise. Low spatial frequency (longer wavelengths) is general structure. If you took out all the high spatial frequency, you would see a blurry image. If you took out all the low frequency information, it would be hard to determine what you're looking at.
Wow, absolutely top notch presentation, Russ! You took a complex set of concepts and did a beautiful job at bringing the audience to that "ah ha" moment where I can clearly see the power of combining Drizzle with Deconvolution, considering the Point Spread Function. A wonderful journey ✨️ and I will definitely apply to my wide field / undersampled / 5-7" per pixel image scale images to improve contrast and resolution 👍 Thank you and keep looking up and enjoying the journey... Cheers!
Fabulous presentation by Russell… i remember signals and systems undergraduate classes not-fondly… from 42 years ago. Russell presented the information in a way that I might have enjoyed those classes back in the early 80’s if I had such a talented instructor who applied the theory to practical uses. Unfortunately I had the graduate student teacher who was terrible and ruined the experience. Luckily semiconductor physics caught me and I had something I could use in my semiconductor focused career! Russell is really upping the game of so many astrophotographers. Thanks Russell!!!
Fantastic presentation. I dont currently use BXT, but will definitely be looking into it. I've recently decided to make the effort to truly understand the functions of the various processing tools, and this presentation has gone a long way to helping me understand deconvolution.
Thanks Russell.
Thanks Russ and TAIC for this informative presentation!
Simply incredible, thank you so much Russ! Using a SharpStar HNT150 (similar to the epsilon) so I can't wait to see the results of the next BXT AI !
That was a terrific presentation leading to several "light bulb" moments. As a relative newcomer to astrophotography and Pixinsight it was invaluable in aiding my understanding of what happens twixt star and pixel
You did a great job presenting deconvolution and hopefully put to bed some of the criticism that "BXT" creates data. You did an even better job in building and training BXT!
I'm using an 11" ~F/1.9 scope at about 1.5 arcsec/sample.. The optics introduce some off-axis aberrations including chromatic.
BXT does a great job at correcting the optical abbs and reducing seeing, most impressive, but nothing compared to what comes next. I reprocessed a couple sessions using 2x drizzle and a .65 spot factor(underwhelming) but I was totally blown away by the BXT output! The low contrast non-stellar improvepment was even more dramatic than the stellar. Detail grew magically (but faithfully) from the mist.
I learned a lot from this presentation, especially on “classical” signal processing. Are there any presentations going into more detail on the neural network / ML side of BXT?
Very admirable explanation of a technically complicated topic. Thanks.
An excellent and very informative presentation. Thank you to TAIC and Russ for this gem!
Beautiful and all and thank you. I bought BlurXt, but I am not impressed by the sudden price increase, from $60 to $100, and from 40 to 90 for the second app. I understand we're living in a cosmic inflationary period but 67% increase seems excessive.
Still probably the best value for money tool available to astrophotographers
Brill!!!!!! Many thanks
Excellent presentation, and job done !! Thanks a lot.
Russ... Great video! Thank you!
Awesome presentation !
another great presentation russ!
excited about what he has achieved.
I have a question Russell ...
is the system good enough to perform photometry?
Is the data lost in the background of the image restored correctly?
Regard
From the BlurXTerminator documentation: Great care has been taken in the design and development of the neural network architecture, and especially in the training methods employed, to try to ensure that BlurXTerminator produces results that are faithful to reality and not "inventive." It is not and never will be perfect in this sense for all images produced by all instruments. Star centroids may not be perfectly maintained. Total flux may not be perfectly conserved. Exact shift-invariance (getting the same result regardless of whether an image is translated by some number of pixels) is not guaranteed. BlurXTerminator's "best guess" at what detailed structure underlies the original blurred image may not be perfectly faithful to reality in all cases.
A consequence of this is that BlurXTerminator should never be used for quantitative or scientific applications. Neural networks and other forms of "machine learning" are indeed finding applications in scientific astronomy, but most often in the form of tools fine-tuned for particular instruments, not general-purpose solutions like BlurXTerminator. Their results are still looked at with a bit of a jaundiced eye, and surprising or potentially important results are always cross-checked using traditional, more transparent methods.
What a refreshing change!
This seems to have led a lot of people to Russ's website--it's been overloaded and is now (temporarily, I hope) offline.
So, spatial frequency is detail?
High spatial frequency is detail (short wavelengths), and very high spatial frequency is generally noise. Low spatial frequency (longer wavelengths) is general structure. If you took out all the high spatial frequency, you would see a blurry image. If you took out all the low frequency information, it would be hard to determine what you're looking at.
Russ is AMAZING!!!!!!!!