Link to the sharpening tool is here: github.com/deepskydetail/AstroSharp Also, I think Blur Exterminator may have 21 million parameters, not 2 million.
Tonight I tried both versions on an Ha stack from last night and the latest version appears to be the clear winner, the first beta version accentuated the noise and appeared to create patterns in it. I look forward to the continued progression of the model, one thing I would suggest is updating the information at the bottom of the program as it still states B&W only. It also might be good to state in there that the image must be stretched as between learning about the tool and using it I had forgotten the details and tried to just go off the information within the program, my mistake but clarity within a program is always nice. Keep up the good work!
Hi! I started a thread on CN. I posted some examples where I first removed the stars on finally processed images and got really good results. Keep up the good work!
Mind-blowing stuff! Just to be clear - where in the post-processing steps would I use this tool if I am using Siril? If I read this correctly, it would be after all the linear operations and with the stretched image - but before further processing in PS or GIMP? My normal workflow in Siril is crop > background extraction > color calibration > noise reduction > deconvolution > stretching > green noise > saturation.. then over to GIMP for some extra curves and noise reduction etc. Seems like i may be able to skip the deconvolution step at least. Experimenting for now, but would be great to get guidance from On High :). THANK YOU for all you do - your SNR video also helped me understand some things I had been missing!
Thanks! The model was trained on stretched images, so I think your process makes sense. Crop, background extraction, stretching, and then use the sharpening tool would make sense to me. I would save noise reduction to the very end because that gets rid of detail. But as long as it works well that's really all that matters.
Looks awesome, I'll need to test this! One thing, you mentioned that the model is trained on stretched images but also that for color images it should be applied before messing with the colors - I assume it would be fine to do do color calibration first though, right? Since it should be done when the image is linear
Thanks! Love your videos btw! Please do test it, and give feedback :) It should be ok to do color calibration linearly and then put it through the tool after stretching. 99% of the time the colors should be the same. The tool assumes the input image is in sRGB/RGB color space, so there could be some conversion issues if it isn't.
I didn’t know about this and I did run on M101 with the new supernova after color calibration and stretch, worked very well but I’m am not a pro or seasoned astrophotographer, but I am a seasoned photoshop user and can judge when color channels get messed up, this wasn’t the case!
Tested this on some of my pictures. The sharpening works nicely, but I noticed that some shadows got crushed, but not to black, but a dark gray. I noticed this on M101, where the thin wisps of the spiral tails disappear during the sharpening process. It's the same with both original and the new model. This means that I actually lose detail in low SNR regions.
Thanks for doing this ! Very interesting stuff . Do you think at some point this could be used on Planetary images or would the AI need to be trained first ? Cheers .
I've downloaded then run the code and selected a colour image of M101 but nothing seems to happen. I can see electron-quick-start running but using very little CPU. The file has been flattened and converted to 16bit. Is there something else required?
I tried it in some of my already processed results. There was some sharpening already applied (deconvolution, wavelets in Siril, or frequency separation in Gimp), and eventhough the tool says not to use sharpened images, the results were stunning. The only image I wasn't very happy with the result was of M13 where it kind of killed the fainter stars around the core. On the other hand I really like what it's doing with the stars on my other images. So thank you very much for the tool.
Nothing is installed on the hard drive. It's a portable app, so all you have to do is delete the folder. You could try reopening the app (sometimes that works) or control+shift+R to force reload it. Sometimes that works.
I would do noise reduction last. NR will get rid of detail that will be hard to get back otherwise. But you could try it out both was to figure out which works best :)
Yes, it is just an Electron app! I'm slowly working to get it to Linux. Tried with Mac but with little success (probably because I was emulating it). I think Linux should be less complicated.
Possibly one day I can make it cross platform. Prior to this, I just didn't have any experience making programs/apps. One possibility is to make the web app work with the current version, and then you could just go online and use it. We'll see though!
Hey, really would love to use this! I've downloaded and extracted but when I get the title/open/splash page it's just blank (white). I get the title menu bar but nothing else. When I go to reload from view nothing happens. Running Windows 11.
@@SamFTurner Thats the exact same behavior I experience intermittently on my new win11 pc. I keep closing and re-opening the program and it finally runs, although it does take about 5 seconds to launch when it does. Also, my logged in account is a member of the local admins group. It seems to happen more frequently when other programs are open in the background, like file explorer, utube and siril for example. When there are no programs open, it happens very rarely if at all. I tried opening/closing the program 10 times (no other programs running) with and without wifi running and it worked perfectly, this AM I just have utube open and it hung first launch and ran the second time..... I hope this give Mark some ammunition to work with in troubleshooting the issue, I thought it was just me.....
Damn what a ghreat update! It's mind blowing. Thanks a lot. Just a question: Is there a way to utilize more CPU Cores or even the GPU CUDA processors for running the network? If not...no biggie... I wait longer for the images from the telescope. hahah....Also it's a good chance get a coffee or so.
Thanks! Right now, there isn't really a way to use more CPU cores or GPU CUDA cores. It could happen with the right code eventually, but I'm trying to prioritize accessibility over speed at the moment.
Thank you so much for your work! I will give it a try myself for sure. Quick question: Might it be beneficial to remove the stars before sharpening? Simply amazing what you are doing for the community! - Stefan
@@deepskydetail Just tested it. Works incredibly well on my oversampled images. Even when I have already applied Deconvolution. Artifacts aren't even really aparent. I will probably end up incorperating your tool in my standard workflow as a final step. - Stefan
I'm looking into it. I currently have virtual box set up with Mac OS on it, and will try to see what I can do. Also, I may be able to update the web application to upload larger files (see previous videos), which would allow anyone to just use it online. Hopefully, one of those two things will make it more accessible.
This is amazing. I wanted to try out Bxt but I didn’t have pix so I haven’t chance. I found out this and just tried it and it works great. I’m wondering if it is possible to go through tool multiple time for better results or deconvolution after process?
Hi, its very impressive that you pulled it off for the color too! I saw that in one of your older videos you said you trained the model using monochrome blurred images and I guess trained the algorithm using color data for implementing sharpening for color images. I don't know if its possible or holds valid or not but how about using Hubble's or JWST's images for training the algorithm?
@@deepskydetail Yes, that is indeed concerning! I wanted to ask if there's a way to implement features like adjusting the star's size present in the frame and reducing the number of stars (star reduction). I'm new to astrophotography! I've been using your software post processing using Photoshop and have obtained better results of the Horsehead Nebula and Orion. When the season arrives, I will gather more data and stack them.
Where is the best place to go for support? Mine doesn't seem to be running in Win11. I pick the file then nothing happens. Am I missing a button to press?
I notice weird behavior on win11 current in updates. At first I 'run as admin' which helped but it would still occationally hang with white background program area display and spinning orange circle but the program does not complete loading or start running. I just keep closing/relaunching and it eventually runs. Do you develop/test running win11?
I've only tested it in Windows 10, but I know that other people have used it on Win 11. Some people have said that the program turns off without an internet connection. Maybe that could be part of the issue? I'm trying to look into it.
If it's totally new to you, I would first check out some RUclips videos on it. 3Blue1Brown has a good playlist on neural networks. Then I'd check out some introductory undergraduate level textbooks on the subject. The textbooks could be related to computer science or even cognitive psychology. You're also going to want to learn how to code in Python (or R; but Python is probably the better choice), and check out the documentation on Keras. Convolution Neural Nets (CNNs) are what's kind of the standard right now for image processing, but mine is really just a simple feedforward model.
Link to the sharpening tool is here: github.com/deepskydetail/AstroSharp
Also, I think Blur Exterminator may have 21 million parameters, not 2 million.
Don't forget to re-pin your pinned comment after editing it (RUclips unpins comments that have been edited) :)
@@CuivTheLazyGeek Thanks! Good to know :)
Incredible how much you can learn in a project like this. Amazing tool!
Thanks!
This is brilliant, tried the mono version. This one is even better. Kudos the hard work.
Thank you!
Great tool! HBM & AC modes always seem to get best results for me. I've tried various settings with DPSF and they tend to result in plenty of noise.
Tonight I tried both versions on an Ha stack from last night and the latest version appears to be the clear winner, the first beta version accentuated the noise and appeared to create patterns in it.
I look forward to the continued progression of the model, one thing I would suggest is updating the information at the bottom of the program as it still states B&W only.
It also might be good to state in there that the image must be stretched as between learning about the tool and using it I had forgotten the details and tried to just go off the information within the program, my mistake but clarity within a program is always nice.
Keep up the good work!
Nice one, Amazing result.
I tested it now on M51 color...Amazing result!!! Many thanks 👏👏👏
Awesome!
Hi! I started a thread on CN. I posted some examples where I first removed the stars on finally processed images and got really good results. Keep up the good work!
Great! Thanks! :)
Nice!!!! I have just downloaded it and will give it a try. Friday was a good night, so got a few interesting targets to try this on.
Great! Let me know how it works :)
Good job on this, I tested it just now and it works really well
I'm glad it works for you!
@@deepskydetail do you plan to add fits support so as to eliminate the switching of file formats in work flow?
@@Hilmi12 Possibly at some point. It's a good idea, but I think getting the model to work better is my main priority.
i just used it, and i am super impressed. it sharpened my image really good but it took a while. but i like it.
Incredible animation technique!
Thank you!
Mind-blowing stuff! Just to be clear - where in the post-processing steps would I use this tool if I am using Siril? If I read this correctly, it would be after all the linear operations and with the stretched image - but before further processing in PS or GIMP? My normal workflow in Siril is crop > background extraction > color calibration > noise reduction > deconvolution > stretching > green noise > saturation.. then over to GIMP for some extra curves and noise reduction etc. Seems like i may be able to skip the deconvolution step at least. Experimenting for now, but would be great to get guidance from On High :). THANK YOU for all you do - your SNR video also helped me understand some things I had been missing!
Thanks! The model was trained on stretched images, so I think your process makes sense. Crop, background extraction, stretching, and then use the sharpening tool would make sense to me. I would save noise reduction to the very end because that gets rid of detail. But as long as it works well that's really all that matters.
This is fantastic! Question though.. Can the First Beta option also be used on a color image, or is that reserved for monochrome?
Both can be used with color images :)
Also I'd just like to say that I enjoy your videos!
@@deepskydetail Thanks! I'm working on a video showcasing your software. I think the community will love it!
Looks awesome, I'll need to test this! One thing, you mentioned that the model is trained on stretched images but also that for color images it should be applied before messing with the colors - I assume it would be fine to do do color calibration first though, right? Since it should be done when the image is linear
Thanks! Love your videos btw! Please do test it, and give feedback :) It should be ok to do color calibration linearly and then put it through the tool after stretching. 99% of the time the colors should be the same. The tool assumes the input image is in sRGB/RGB color space, so there could be some conversion issues if it isn't.
I didn’t know about this and I did run on M101 with the new supernova after color calibration and stretch, worked very well but I’m am not a pro or seasoned astrophotographer, but I am a seasoned photoshop user and can judge when color channels get messed up, this wasn’t the case!
Tested this on some of my pictures. The sharpening works nicely, but I noticed that some shadows got crushed, but not to black, but a dark gray. I noticed this on M101, where the thin wisps of the spiral tails disappear during the sharpening process. It's the same with both original and the new model. This means that I actually lose detail in low SNR regions.
Thanks for the feedback. This is helpful! Do you mind sharing your setup?
@@deepskydetail I'll email you some comparison images soon
Great thanks!
@@deepskydetail Email sent. I hope it finds you well. Let me know if you need anything else.
Nice job, I think I'll run the two version and use lighten in photoshop to merge them
The extraction of the ZIP takes forever. So I have enough time to mention this here. I'm looking forward to check the tool out anyway.
Thanks for doing this ! Very interesting stuff . Do you think at some point this could be used on Planetary images or would the AI need to be trained first ? Cheers .
It would probably require another ai tool trained for that purpose. But you could give it a try and see how it does :)
Looks awesome. Are you planning to monetize it? Or will you make it open source?
I have no plans of monetizing it at the moment :)
Just blows my mind 🎉
I've downloaded then run the code and selected a colour image of M101 but nothing seems to happen. I can see electron-quick-start running but using very little CPU. The file has been flattened and converted to 16bit. Is there something else required?
Solved. I'd put the code in a read only folder. It needs to write files in its folder.
Glad you got it working! I was just about to ask whether you could see the progress bar.
I tried it in some of my already processed results. There was some sharpening already applied (deconvolution, wavelets in Siril, or frequency separation in Gimp), and eventhough the tool says not to use sharpened images, the results were stunning. The only image I wasn't very happy with the result was of M13 where it kind of killed the fainter stars around the core. On the other hand I really like what it's doing with the stars on my other images. So thank you very much for the tool.
I'm glad you liked it! If you would like to send the results of M13 to my email, feel free to :)
great job, great tool, great idea!!
does something similar for noise reduction exist??
There is something called AstroDenoise, but it seems a bit harder to install. I haven't tried it, but it's worth checking out!
Большое спасибо. Великолепная работа!
Пожалуйста... и спасибо тоже!
thank you for this free tool
I think my installation went ary as its not working. Everything is just greyed out. How do I uninstall cleanly before reinstalling?
Nothing is installed on the hard drive. It's a portable app, so all you have to do is delete the folder. You could try reopening the app (sometimes that works) or control+shift+R to force reload it. Sometimes that works.
Do you recommend running the noise reduction tools in siril before running your image through astrosharp? Thx...
I would do noise reduction last. NR will get rid of detail that will be hard to get back otherwise. But you could try it out both was to figure out which works best :)
@@deepskydetail Thx...
Great work! I believe this is an Electron app (not sure), could you plan on porting it to Linux in the near future?
Yes, it is just an Electron app! I'm slowly working to get it to Linux. Tried with Mac but with little success (probably because I was emulating it). I think Linux should be less complicated.
Fun to see the progress! Tyvm for doing this! Any plans for cross platform? Any specific library holding you back?
Possibly one day I can make it cross platform. Prior to this, I just didn't have any experience making programs/apps. One possibility is to make the web app work with the current version, and then you could just go online and use it. We'll see though!
Hey, really would love to use this! I've downloaded and extracted but when I get the title/open/splash page it's just blank (white). I get the title menu bar but nothing else. When I go to reload from view nothing happens. Running Windows 11.
Do you also see the spinning orange circle? Try running the program as admin. Rt. Click executable and select 'run as admin'
@@bonzer2u spinning orange circle seen. Run as administrator, no dice still.
@@SamFTurner Thats the exact same behavior I experience intermittently on my new win11 pc. I keep closing and re-opening the program and it finally runs, although it does take about 5 seconds to launch when it does. Also, my logged in account is a member of the local admins group. It seems to happen more frequently when other programs are open in the background, like file explorer, utube and siril for example. When there are no programs open, it happens very rarely if at all. I tried opening/closing the program 10 times (no other programs running) with and without wifi running and it worked perfectly, this AM I just have utube open and it hung first launch and ran the second time.....
I hope this give Mark some ammunition to work with in troubleshooting the issue, I thought it was just me.....
Thanks for the feedback. I'll try to figure out what is going on. Some have found that toggling their internet on and off has helped for some reason.
Damn what a ghreat update! It's mind blowing. Thanks a lot. Just a question: Is there a way to utilize more CPU Cores or even the GPU CUDA processors for running the network? If not...no biggie... I wait longer for the images from the telescope. hahah....Also it's a good chance get a coffee or so.
Thanks! Right now, there isn't really a way to use more CPU cores or GPU CUDA cores. It could happen with the right code eventually, but I'm trying to prioritize accessibility over speed at the moment.
Thank you so much for your work! I will give it a try myself for sure.
Quick question: Might it be beneficial to remove the stars before sharpening?
Simply amazing what you are doing for the community!
- Stefan
You can remove stars before sharpening if you'd like. That is an option that could work :)
@@deepskydetail Just tested it. Works incredibly well on my oversampled images. Even when I have already applied Deconvolution. Artifacts aren't even really aparent.
I will probably end up incorperating your tool in my standard workflow as a final step.
- Stefan
I’m thrilled to see this! But unfortunately I’m a Mac user with no access to Windows machines.
Is there a chance of a MacOS port of this in future? 🤞🥺
I'm looking into it. I currently have virtual box set up with Mac OS on it, and will try to see what I can do.
Also, I may be able to update the web application to upload larger files (see previous videos), which would allow anyone to just use it online.
Hopefully, one of those two things will make it more accessible.
@@deepskydetail Oooh, glad to hear it! Definitely keep us informed please!
Is here on your RUclips the best way to look for updates?
Yes, probably for the time being :)
You got it!
Nickel 👌…merci à toi…!!!
You're welcome and thanks for watching!
This is amazing. I wanted to try out Bxt but I didn’t have pix so I haven’t chance. I found out this and just tried it and it works great. I’m wondering if it is possible to go through tool multiple time for better results or deconvolution after process?
You could try to do that and see if it works better. It might over sharpen some of the details though. But it's worth a shot!
Does this work on Planets?
Some people have reported that it works ok with planets and lunar images. It was trained on DSOs. As long as it's in tiff format you can try it :)
Any interest in making it open source? You could collaborate with the devs from Siril.
I think that is a good idea, but I'd like to get the model a bit more polished before it goes to Siril.
Only Windows??? :) Nice work man! Looking forward for the Apple version at some point :)
Hi, its very impressive that you pulled it off for the color too! I saw that in one of your older videos you said you trained the model using monochrome blurred images and I guess trained the algorithm using color data for implementing sharpening for color images. I don't know if its possible or holds valid or not but how about using Hubble's or JWST's images for training the algorithm?
Thanks! The model is trained on hubble images. I might include JWST too, but I'm worried about the star diffraction patterns. So maybe :)
@@deepskydetail Yes, that is indeed concerning! I wanted to ask if there's a way to implement features like adjusting the star's size present in the frame and reducing the number of stars (star reduction). I'm new to astrophotography! I've been using your software post processing using Photoshop and have obtained better results of the Horsehead Nebula and Orion. When the season arrives, I will gather more data and stack them.
Where is the best place to go for support? Mine doesn't seem to be running in Win11. I pick the file then nothing happens. Am I missing a button to press?
Do you mean the program won't start, or when you browse for the file and select it the progress bar at the bottom doesn't show up after it is loaded?
@@deepskydetail The progress bar doesn't ever show up.
I notice weird behavior on win11 current in updates. At first I 'run as admin' which helped but it would still occationally hang with white background program area display and spinning orange circle but the program does not complete loading or start running. I just keep closing/relaunching and it eventually runs. Do you develop/test running win11?
Note: my account is a member of the local admins group.....
I've only tested it in Windows 10, but I know that other people have used it on Win 11. Some people have said that the program turns off without an internet connection. Maybe that could be part of the issue? I'm trying to look into it.
I am interested in learning how to build more ai image tools like this. Any recomended reading on where to get started?
If it's totally new to you, I would first check out some RUclips videos on it. 3Blue1Brown has a good playlist on neural networks. Then I'd check out some introductory undergraduate level textbooks on the subject. The textbooks could be related to computer science or even cognitive psychology. You're also going to want to learn how to code in Python (or R; but Python is probably the better choice), and check out the documentation on Keras. Convolution Neural Nets (CNNs) are what's kind of the standard right now for image processing, but mine is really just a simple feedforward model.
@@deepskydetail thanks. I do old school image processing in Matlab and just need some projects to jump over to python.
does this need internet
No, it shouldn't need the internet.
Thxxxxxxxx!!!👏👏👏👏
👍💪