VideoHelp Forum
+ Reply to Thread
Results 1 to 23 of 23
Thread
  1. Sigh... what the actual **** is it with these impuritans that keep releasing obviously-upscaled shit on Blu-ray? Only maybe 20% of the scenes are at least 720p detail, the rest look upscaled even when resized to 400p.

    Is there a video-equivalent of those hideously slow deblur tools for photoshop? Save your lectures about this being impossible, I'll be happy to even get it to 480p crispness so it shouldn't be too unrealistic.
    Quote Quote  
  2. Banned
    Join Date: Oct 2004
    Location: New York, US
    Search Comp PM
    -30-
    Last edited by sanlyn; 23rd Mar 2014 at 13:40.
    Quote Quote  
  3. Member
    Join Date: Sep 2007
    Location: Canada
    Search Comp PM
    Typical low quality blu-ray animation releases from cheap studios. Lots of upscaled SD released these days

    Photoshop's filter doesn't work for those types of situation. It's essentially a deconvolution filter, so if you have something like a directional blur , that sort of thing where vectors can be calculated.
    Quote Quote  
  4. This is not an animated movie and there is no directional blur, just simple blur. I think it's sourced from a faulty master supervised by moron film technicians rather than a DVD that was upscaled, though, because there are scenes that are HD sometimes. The only true 1080p detail is in scenes containing text like credits.

    I think the filming crew used a too wide lens which they later cropped and zoomed for the close-up scenes, hence the appearance of blown-up frames. I'm not sure.

    Anyway, let me illustrate: I'm looking for any video-equivalent of a filter that can do this. I know it looks like shit but its way better when resized to half its resolution and I only intend it to look good at 480p anyway.
    Attached Files
    Quote Quote  
  5. LimitedSharpen, MSharpen, asharpen, SeeSaw, WarpSharp,.. there is a ton of sharpen filters out there, that might do the job,...
    Quote Quote  
  6. Is sharpen the same as deconvolution?
    Quote Quote  
  7. not necessarily, deconvolution can be used as a sharpening technique,...
    Quote Quote  
  8. Member
    Join Date: Mar 2011
    Location: Nova Scotia, Canada
    Search Comp PM
    I've actually found that the best way to deal with that sort of crud usually is to use the dithering filter in SMPlayer. "Add noise" on the video filters menu. Especially with pixelation.

    Sounds ass backwards but dithering is routine in digital audio (well, you shouldn't need it with 24 bit). It masks quantizing distortion. It doesn't really sound right but it works.

    I wouldn't expect all BD or DVD transfers to be great nowadays. Not with sales so flat. First chance they get to move all video and audio to the cloud they'll do it, and physical media will go the way of betamax.
    Quote Quote  
  9. Member 2Bdecided's Avatar
    Join Date: Nov 2007
    Location: United Kingdom
    Search Comp PM
    It would help if you posted a clip.

    This is another AVIsynth tool aimed at what you want...
    http://avisynth.org/vcmohan/FQSharp/FQSharp.html

    Another approach is to downscale to the "real" resolution, and then use the various upscaling tricks (nnedi3, limitedsharpen, adding grain, etc) to fake a higher resolution.

    I've never seen an improvement that made me go "wow" on real content. The impressive comparisons tend to use content where the blur function is known - that's not a real life situation.

    Originally Posted by Hoser Rob View Post
    I've actually found that the best way to deal with that sort of crud usually is to use the dithering filter in SMPlayer. "Add noise" on the video filters menu. Especially with pixelation.

    Sounds ass backwards but dithering is routine in digital audio (well, you shouldn't need it with 24 bit). It masks quantizing distortion.
    Adding dither before you quantise (or reduce the number of bits) prevents quantisation distortion and maintains real amplitude information below the least significant bit. This works the same in audio and video. Adding noise to a blurry image is just tricking the eye into thinking it sees details - details that aren't real. The two aren't really comparable (though I use both!).

    I wouldn't expect all BD or DVD transfers to be great nowadays. Not with sales so flat. First chance they get to move all video and audio to the cloud they'll do it, and physical media will go the way of betamax.
    hello future

    Cheers,
    David.
    Quote Quote  
  10. Member
    Join Date: Jul 2008
    Location: Old Dominion
    Search Comp PM
    Originally Posted by Hoser Rob View Post
    I wouldn't expect all BD or DVD transfers to be great nowadays. Not with sales so flat. First chance they get to move all video and audio to the cloud they'll do it, and physical media will go the way of betamax.
    Sadly moving to "the cloud" seems like more of a media control mechanism than a technological leap forward.
    Quote Quote  
  11. I tried FQRestore and it sharpens rather than deblurs. The test on that blurry I image I posted was worse quality than what I got with the unknown photoshop tool, but as I said, even if I could get results as good as I can on PS then it's all good.

    FGRestore doesnt seem capable of this, unless I'm using it wrong. Do I have to do FQVisual before FQRestore? The only real setting I see on FGRestore is the x= and y=. Anything above 2 looks like shit. Other settings don't significantly change filtration. The scale setting seems like a contrast/brightness setting.

    I've never seen an improvement that made me go "wow" on real content. The impressive comparisons tend to use content where the blur function is known - that's not a real life situation.
    True, but I found this impressive enough:
    Attached Images    
    Quote Quote  
  12. Member
    Join Date: Sep 2007
    Location: Canada
    Search Comp PM
    But you're not doing this for forensic recovery? Presumably you're trying to make it look sharper and pleasant to watch. These are entirely different goals.

    Like you said, the photoshop result in the 1st zip file "look like shit" - it's very easy oversharpen and it's not pleasant to watch on real video content. In motion , the halos/ringing and aliased edges will vibrate and look very bad . Everything is unnatural looking with warping in the wrong direction (some objects are actually made worse because of this)

    You're better off using dedicated video sharpeners, +/- grain
    Quote Quote  
  13. It looks bad on the full 1080p frame but much better when scaled down twice as smaller. The result with FQRestore was much worse than my photoshop result, that's what I'm saying. There were way more halos and ringing artifacts and this was on the lowest possible setting of x=2 and y=2. All the other options changed nothing except brightness. It also didn't look like it made any attempt to deconvolute.

    No I'm not doing this for forensic recovery, but I am trying to restore enough details so it at least doesn't look blurred on 480p, because I know anything higher will look artificial/artifacted.

    What dedicated tools do you recommend?
    Quote Quote  
  14. Member
    Join Date: Sep 2007
    Location: Canada
    Search Comp PM
    Do you know what to look for? Dedicated video sharpeners will oversharpen if settings are too strong (oversharpen halo artifacts), but they are natural looking - edges retain their natural relationships. They won't give you the typical deconvolution artifacts which make everything look artificial (they look like halos but much worse, often you will get "ghosting" artifacts as object edges are interpolated incorrectly - the directional blur isn't calculated correctly)

    Here is extreme example comparing them @ 1280x720 , and pointing out the deconvolution errors. You normally wouldn't sharpen this much using avisynth (halos too, but they are more natural looking) , it's just done that much for comparison purposes. Grain is also added, to simulate the perception of texture (for example, compare the hair , pants, wood counter, etc...)
    Attached Files
    Last edited by poisondeathray; 30th Apr 2012 at 15:07.
    Quote Quote  
  15. Originally Posted by poisondeathray View Post
    Do you know what to look for? Dedicated video sharpeners will oversharpen if settings are too strong (oversharpen halo artifacts), but they are natural looking - edges retain their natural relationships. They won't give you the typical deconvolution artifacts which make everything look artificial (they look like halos but much worse, often you will get "ghosting" artifacts as object edges are interpolated incorrectly - the directional blur isn't calculated correctly)

    Here is extreme example comparing them @ 1280x720 , and pointing out the deconvolution errors. You normally wouldn't sharpen this much using avisynth (halos too, but they are more natural looking) , it's just done that much for comparison purposes. Grain is also added, to simulate the perception of texture (for example, compare the hair , pants, wood counter, etc...)
    Dayum! That is impressive! How did you do that?

    And no, I don't know what to look for. I know it's not directional blur, but I don't know what type of blur this is because I can't figure out if this is an upscaled DVD or just a blurry overzoomed master. If this was regular lens blur, I assume halos would not be ran into because the radius would be constant (or would it?)

    Artificially-produced blur can be undone with 99% crispness but real-life blur such as the "Read Me" example I posted above that was taken with an unfocused camera could not be.
    Quote Quote  
  16. Member
    Join Date: Sep 2007
    Location: Canada
    Search Comp PM
    Originally Posted by Mephesto View Post

    Dayum! That is impressive! How did you do that?
    Not that impressive IMO: it's oversharpened on purpose to match the oversharpened version you provided. Watch an actual video with those adjustments and you will see what I mean. You're only looking at still frame. You can get away with a lot more on still images. "Normal" video sharpeners usually have limiting provisions so you don't oversharpen so much

    And no, I don't know what to look for. I know it's not directional blur, but I don't know what type of blur this is because I can't figure out if this is an upscaled DVD or just a blurry overzoomed master. If this was regular lens blur, I assume halos would not be ran into because the radius would be constant (or would it?)
    I meant what artifacts/problems to look for in terms of using deconvolution "deblur tool" for video . Those types of artifacts have a sort of "signature" if you will. Did you compare them by flipping back & forth? Notice the problem areas, like edge outlines, things like specular highlights are accentuated the wrong direction? Look at the "American" text, notice the ringing pattern; halos are worse overall. They both have problems, but I would say the deconvolution approach looks worse, because the wrong data was dialed in. Deconvolution approach usually only work well when you know the specific parameters (or they can be calculated) like motion vectors, lens parameters , focal length, shutter speed, diffraction coefficients, pan speed etc... ie. not that useful in real world video cases. Did you know Adobe admitted the Photoshop demo a few months back was actually faked? It's the same crap, different pile, as software like focus magic, topaz infocus, etc...

    "Regular" lens blur would have a radial effect where there will be concentric gradations in the magnitude of blur, but the degree of blur can shift as the DP focuses/defocuses on subjects, especially with shallow DoF shots, so you really have to examine on video, not a still. Your example was blurry overall so it's very unlikely to be the case of "regular lens blur" - which can be "undone" to an extent if you know the specific parameters

    Artificially-produced blur can be undone with 99% crispness but real-life blur such as the "Read Me" example I posted above
    that was taken with an unfocused camera could not be.
    I don't know about "99%", but it depends on many things including if you have access to the method and data to how it was actually blurred



    This is the script I used for this image example (but again, you wouldn't use this on real video , or you'd use a lot weaker modified settings. You wouldn't need to ConvertToYV24 or YV12 either).

    Here is a list of sharpeners you could try (there are many more)
    http://avisynth.org/mediawiki/External_filters#Sharpeners

    Some of the more exotic ones like SSSharp , MedSharp , are very slow. Didee also has a few , this one is show in many classic examples, but it doesn't even have a proper name (http://forum.gleitz.info/showthread.php?t=33835&p=383875)


    Code:
    ImageSource("blur.jpg")
    ConvertToYV24()
    mt_lutxy(binomialblur(1.0),yexpr="x x y - abs 1 x y - 10 / / atan * +",u=2,v=2)
    Spline36Resize(1280,720)
    ConvertToYV12()
    mt_lutxy(binomialblur(1.0),yexpr="x x y - abs 1 x y - 10 / / atan * +",u=2,v=2)
    GrainFactory()
    Quote Quote  
  17. Member 2Bdecided's Avatar
    Join Date: Nov 2007
    Location: United Kingdom
    Search Comp PM
    Originally Posted by poisondeathray View Post
    Here is extreme example comparing them @ 1280x720 , and pointing out the deconvolution errors. You normally wouldn't sharpen this much using avisynth (halos too, but they are more natural looking) , it's just done that much for comparison purposes.
    Lots of TVs will sharpen almost this much by default - and more than this if you mess with the settings.

    People who care about video quality tend to turn the sharpness control down quite a lot.

    AVIsynth seesaw, limitedsharpenfaster, and sssharp (super slow sharpen) are all useful with half-decent sources. It you go overboard, not only does the result look unnatural, but it becomes very difficult to encode.

    Cheers,
    David.
    Quote Quote  
  18. Originally Posted by poisondeathray View Post
    Not that impressive IMO: it's oversharpened on purpose to match the oversharpened version you provided. Watch an actual video with those adjustments and you will see what I mean. You're only looking at still frame. You can get away with a lot more on still images. "Normal" video sharpeners usually have limiting provisions so you don't oversharpen so much
    Granted, I didn't test it on a video yet but it looks fine to me. Not oversharpened at all. No halos.

    Originally Posted by poisondeathray View Post
    I meant what artifacts/problems to look for in terms of using deconvolution "deblur tool" for video . Those types of artifacts have a sort of "signature" if you will. Did you compare them by flipping back & forth? Notice the problem areas, like edge outlines, things like specular highlights are accentuated the wrong direction? Look at the "American" text, notice the ringing pattern; halos are worse overall. They both have problems, but I would say the deconvolution approach looks worse, because the wrong data was dialed in. Deconvolution approach usually only work well when you know the specific parameters (or they can be calculated) like motion vectors, lens parameters , focal length, shutter speed, diffraction coefficients, pan speed etc... ie. not that useful in real world video cases.
    Well, I didn't test them on an actual video (I didn't know of any video deconvolution tools until now) so even if the occasional halos and ringing isn't too bad on the picture, I imagine it's a mess on the video. I wish I knew more about deconvolution and see how good I can restore that particular screenshot. I know about lens parameters, but what is focal length and diffraction? I always thought lens was the most important parameter to know but I don't know how to properly test it, and even then you said there's a lot more to consider.

    Did you know Adobe admitted the Photoshop demo a few months back was actually faked? It's the same crap, different pile, as software like focus magic, topaz infocus, etc...
    Hah, I read about this a couple days before your post. I was impressed (well, wishful rather) by the presentation but several things were off-key and the candidates seemed so amateurish. Then they showed that third example of that dog and how the unblur algorithm can't work because of "not enough edges" yet there were plenty obvious edges in the photo and the before-after shots were identical despite being "processed". The whole thing sounded so weird, and then finally I found the reports where adobe admitted to cherry-picking their photos from 3000 that they took of the same spot mounted by an expensive tripod. Not exactly something you or me would do when taking an average photo with a Parkinsons-afflicted hand.

    What I wonder is, why would Adobe pull shit like this? Does it not harm their credibility? Adobe has brought on revolutionary technologies like that content-aware fill and others. Why minimize that?

    "Regular" lens blur would have a radial effect where there will be concentric gradations in the magnitude of blur, but the degree of blur can shift as the DP focuses/defocuses on subjects, especially with shallow DoF shots, so you really have to examine on video, not a still. Your example was blurry overall so it's very unlikely to be the case of "regular lens blur" - which can be "undone" to an extent if you know the specific parameters
    It was blurry overall, that's what I meant. I assumed this was regular lens blur. Keep in mind this is straight from the Blu-ray. Not all scenes are like this, though. Some are very detailed, some blurry like this one. Most are worse.

    I don't know about "99%", but it depends on many things including if you have access to the method and data to how it was actually blurred
    That's what I mean, yes.

    Some of the more exotic ones like SSSharp , MedSharp , are very slow. Didee also has a few , this one is show in many classic examples, but it doesn't even have a proper name (http://forum.gleitz.info/showthread.php?t=33835&p=383875)
    Couldn't google-translate that page, it wouldn't let me navigate the site. How did you find this? Do I have a Deutschman in my thread?

    ImageSource("blur.jpg")
    ConvertToYV24()
    mt_lutxy(binomialblur(1.0),yexpr="x x y - abs 1 x y - 10 / / atan * +",u=2,v=2)
    Spline36Resize(1280,720)
    ConvertToYV12()
    mt_lutxy(binomialblur(1.0),yexpr="x x y - abs 1 x y - 10 / / atan * +",u=2,v=2)
    GrainFactory()
    Can you explain the parameters and why you used the ones you used? None made significant difference except the 1.0 when I increase it. Why did you use the svript twice?

    Also, I don't have GrainFactory and could only find GrainFactory3 on the internet. Can you upload yours? Not sure I'd actually use it though. I can add grain with ffdshow upon playback
    Quote Quote  
  19. Member
    Join Date: Sep 2007
    Location: Canada
    Search Comp PM
    Originally Posted by Mephesto View Post

    Some of the more exotic ones like SSSharp , MedSharp , are very slow. Didee also has a few , this one is show in many classic examples, but it doesn't even have a proper name (http://forum.gleitz.info/showthread.php?t=33835&p=383875)
    Couldn't google-translate that page, it wouldn't let me navigate the site. How did you find this? Do I have a Deutschman in my thread?
    It was a link to a link to a link somewhere on Doom9. If you notice the author, Didee is one of the master gurus behind many, many functions. Sometimes he comes down from the mountain and writes a function or script. The problem is many of his gems are scattered. I have a bunch bookmarked. It's amazing what he comes up with

    ImageSource("blur.jpg")
    ConvertToYV24()
    mt_lutxy(binomialblur(1.0),yexpr="x x y - abs 1 x y - 10 / / atan * +",u=2,v=2)
    Spline36Resize(1280,720)
    ConvertToYV12()
    mt_lutxy(binomialblur(1.0),yexpr="x x y - abs 1 x y - 10 / / atan * +",u=2,v=2)
    GrainFactory()
    Can you explain the parameters and why you used the ones you used? None made significant difference except the 1.0 when I increase it.
    Why did you use the svript twice?
    Yes the only parameter is to change the binomialblur, higher value = stronger sharpen as you noticed. I used it twice because I would trying to roughly match the "sharpness" to the photoshop pic, just to show that you can get something similar (but better without the deconvolution artifacts). But supposedly there is a science behind multistage sharpening - thats what some of the other advanced filters do like SSSharp, but I don't know the math or the science, you have to ask one of the gurus


    Also, I don't have GrainFactory and could only find GrainFactory3 on the internet. Can you upload yours? Not sure I'd actually use it though. I can add grain with ffdshow upon playback
    It's probably the same. It's grainfactory3.avsi, but the call is GrainFactory() . Note the script has hardcoded values that are not adjustable - when tweaking you actually have to change the script itself (e.g. grain sharpness & strength for particle 1 &2 are adjustable, but things like size, texture need to be changed within the function, at least in the version I have)
    Last edited by poisondeathray; 3rd May 2012 at 20:11.
    Quote Quote  
  20. Originally Posted by poisondeathray View Post
    Yes the only parameter is to change the binomialblur, higher value = stronger sharpen as you noticed. I used it twice because I would trying to roughly match the "sharpness" to the photoshop pic, just to show that you can get something similar (but better without the deconvolution artifacts). But supposedly there is a science behind multistage sharpening - thats what some of the other advanced filters do like SSSharp, but I don't know the math or the science, you have to ask one of the gurus
    I tried SSSharp and a few others and the results were a bit worse than your fast binomialblur script. Maybe I came across the wrong one.

    I'm sure there are many elites on Doom9 but I can't stand even looking at that site let alone going out my way to make a thread on it. The bizarre ******ronics of the top, well-renowned crew completely alienated any trust and respect I would've otherwise had. When the best of the best act like colossal mental ****s on that site, you don't even wanna waste time finding out what the underlings are like.

    It's probably the same. It's grainfactory3.avsi, but the call is GrainFactory() . Note the script has hardcoded values that are not adjustable - when tweaking you actually have to change the script itself (e.g. grain sharpness & strength for particle 1 &2 are adjustable, but things like size, texture need to be changed within the function, at least in the version I have)
    It's not the same, putting just grainfactory() brings up an error. It doesn't matter though, burning grain into the video is counterproductive when it can be done in post-processing.

    Thanks for all the help. It made the video much clearer and detailed. The deconvolution engine of that script couldn't really pull off a phenomenal stunt to bring back 720p detail but as I gave my word, I'm happy to get it to look crisp on 480p. It looks way better than the DVD equivalent.

    Deblurring in general is a hot topic of interest for me and I have several photos that I've kept stored for a long time hoping to someday fix them. I understand that to deblur a real image taken with a real camera, the precise parameters must be known.

    Is there a way to find out by trial and error of the radius blur? What is a good program/tool/plugin for this? The best one that I have is Cathexis and it is seriously slow. It takes 10-20 seconds just for a 150x100 image and there are too many parameters to set.

    I'm also aware that noise is the biggest barrier to deblurring. Since there are no high-freq details to lose, can the image be denoised by conventional means or will this harm the delocalized detail somehow?
    Quote Quote  
  21. Member
    Join Date: Sep 2007
    Location: Canada
    Search Comp PM
    Originally Posted by Mephesto View Post

    It's probably the same. It's grainfactory3.avsi, but the call is GrainFactory() . Note the script has hardcoded values that are not adjustable - when tweaking you actually have to change the script itself (e.g. grain sharpness & strength for particle 1 &2 are adjustable, but things like size, texture need to be changed within the function, at least in the version I have)
    It's not the same, putting just grainfactory() brings up an error. It doesn't matter though, burning grain into the video is counterproductive when it can be done in post-processing.
    Attached.

    Not necessarily counterproductive at all - it depends what your trying to do - virtually all studios do it.

    By your rationale, you could sharpen in playback post processing as well (sharpening is counterproductive to compression as well)



    Is there a way to find out by trial and error of the radius blur? What is a good program/tool/plugin for this? The best one that I have is Cathexis and it is seriously slow. It takes 10-20 seconds just for a 150x100 image and there are too many parameters to set.
    That's just it - trial and error. I don't know any good programs (I've tried a bunch, beta tested a few as well) - they all fail or have severe limitations for real world usage. And the results still aren't "pleasant". Certainly details are more clear, but so are deconvolution artifacts

    For single photo it's might be reasonable to waste few hours on a single photo, but when you do video, forget about it. Consider something like a whip pan - each frame has slightly different characteristics - the velocity changes over that timespan as you accelerate/decelerate. You cannot reuse many of the previously calculated parameters, so you have to recalculate almost everything from scratch. Add in lens distortions, rolling shutter on todays CMOS cameras, focus shifts ...

    I'm also aware that noise is the biggest barrier to deblurring. Since there are no high-freq details to lose, can the image be denoised by conventional means or will this harm the delocalized detail somehow?
    Denoising is always harmful - there is no way you can accurately remove only noise. You will always remove some signal as well. So it becomes a balancing act. You end up doing many iterations of denoise, dering, adjust the many parameters, going forwards, backwards, rinse, repeat.
    Attached Files
    Quote Quote  
  22. Member
    Join Date: Aug 2007
    Location: South Africa
    Search Comp PM
    @Mephesto,

    I just came across this interesting thread. How are you doing with this? Any other exciting new tools in this respect you've come across?

    If it isn't too late already, you might also consider this script by Didée that poisondeathray pointed me to: http://forum.doom9.org/showthread.php?p=1595531#post1595531. It uses topological reconstruction, which I think is probably the best approach of the lot for general uniform blurring like this.
    Quote Quote  
  23. fvisagie, I looked at the screenshots and all I see is selective edge sharpening, no deblurring. Not impressed.
    The most impressive image deblurring tool I've come across is Cathexis but it is way too slow and there are too many choices. Using this on an HD movie is suicide.

    Not necessarily counterproductive at all - it depends what your trying to do - virtually all studios do it.
    My goal is always the same: lowest possible bitrate and best possible quality. Noise ****s that up real good.

    By your rationale, you could sharpen in playback post processing as well (sharpening is counterproductive to compression as well)
    But you see I resize down to the real resolution if I have a blurry video and thus sharpening would be useless because there's no blur. This way, I got the detail up from 320p to about 480p thanks to your tool. Besides, most newbs have no idea how to post-process. For a start, I never do it because I would have to keep changing the parameters for different movies, which is annoying. There should be a universal standard for all encodes before they are released so people don't have to waste time screwing around with their playback settings and most (non-retarded) people I've talked to prefer a pristine image to "grainy warmness" whatever the hell that means.

    Denoising is always harmful - there is no way you can accurately remove only noise.
    I mean the immediate noise before deconvolution. Unless there is delocalized noise in the image, which I don't see is possible.
    Quote Quote  



Similar Threads