logo Sign In

Star Wars GOUT in HD using super resolution algorithm (* unfinished project *) — Page 3

Author
Time
 (Edited)

Two more comparisons for precise bicubic (A = -0.75) vs super resolution. In this case both sources are uncompressed. In both cases details are more clear in the super resolution version. The grain also seems to be somewhat more visible though, and a number of edge artifacts are clearly visible in the second screenshot, which is of course very undesirable.

http://screenshotcomparison.com/comparison/118200

http://screenshotcomparison.com/comparison/118201

Author
Time
 (Edited)

Another comparison (two video samples) between precise bicubic (A = -0.75) and the super resolution. 

Update: Clips removed to make space for new clips.

Author
Time

The BasiClean+MagicUp has less grain, but also considerably less detail, visible in Obi-Wan's beard and Luke's hair, but especially in the string in the left corner of the frame on Obi-Wan's hood. 

Author
Time

Forgot to mention I added also a grain plate...

Agree, less small details, but more precise borders and less artefacts... for example, look at Obi Uan's cap profile... it's full of "jagged lines"... different methods, different results, subjective quality! (^^,)

Sadly my projects are lost due to an HDD crash… 😦 | [Fundamental Collection] thread | blog.spoRv.com | fan preservation forum: fanres.com

Author
Time

Super Resolution only exists in papers of shortsighted academics who have a poor grasp on sampling theory. 

-G

Author
Time

There's only so much detail you can get out of the GOUT.

Not only does the PAL version (of SW and ESB) actually have less detail in it than the NTSC but it has the same problems, which is a weird kind of horizontal line aliasing that looks as if somebody removed one of the fields in higher resolution and then scaled it down. (I tried recreating this using Harmy's version once and the aliasing ended up identical)

The DVNR smearing from the 1993 LD master makes it even worse.

It's as if everything that could possibly be done to prevent a really good looking upscale of this version was done to it.

When I worked on Team Blu's upscale method for the GOUT, I simply tried to keep as much detail as possible while still keeping the "sharp" look but reduce the aliasing. Some may think our version ended up too sharp but if you had seen what it looked like without the sharpening, you would've agreed the sharpening made it look closer to the source.

That said, here's a comparison of the Precise bicubic and Team Blu's upscale method. (Without our color correction)

http://screenshotcomparison.com/comparison/118261

Author
Time

Great results, You_Too!

A question: did you use the NTSC GOUT DVD as source?

Sadly my projects are lost due to an HDD crash… 😦 | [Fundamental Collection] thread | blog.spoRv.com | fan preservation forum: fanres.com

Author
Time

DrDre said:

I think one 2014 Nobel prize winner would disagree:

http://www.nobelprize.org/nobel_prizes/chemistry/laureates/2014/advanced-chemistryprize2014.pdf

 Keeping in mind that I'm a biologist and we sometimes use these techniques (TIRF, STED, STORM) but don't really properly understand them the way a physicist would:

I really don't follow how you're suggesting these specific techniques used for improving the resolution of fluorescence microscopy are related to enhancing images with a fixed number of pixels. Can you please explain? And to what technique specifically are you comparing it to? Temporal tricks like STORM, but across multiple frames rather than relying on the 'flickering' of fluorophores?

Author
Time

_,,,^..^,,,_ said:

Great results, You_Too!

A question: did you use the NTSC GOUT DVD as source?

Thanks. It's the same as the version we released already, except I removed the color correction stuff from the script to be able to compare it to this thread's new algorithm.

Yes, since the NTSC DVD has more detail in it, (and I have tried upscaling both in the past) it's the best source for an upscale.

By the way, I hope nobody thinks I try to advertise anything here. I'm just saying that maybe the super resolution algorithm works better on a source lacking all the downsides the GOUT has. It would be interesting to try it on something that already looks good, some very well mastered DVD release. (Avatar maybe? Wasn't that one considered one of the best?) And then comparing it to a simple Spline64Resize in avisynth.

Author
Time

You_Too said:


It would be interesting to try it on something that already looks good, some very well mastered DVD release. (Avatar maybe? Wasn't that one considered one of the best?) And then comparing it to a simple Spline64Resize in avisynth.

Yes, we should start an "upscaling war"! (^^,)

Sadly my projects are lost due to an HDD crash… 😦 | [Fundamental Collection] thread | blog.spoRv.com | fan preservation forum: fanres.com

Author
Time
 (Edited)

The basic concept for super resolution used for fluorescence microscopy and digital sensors is the same, namely combining multiple low resolution images to obtain a high resolution image. In the first case it is called optical super resolution. In the second case it is called geometrical super resolution. Since a video is nothing more than a sequence of very similar images, the concept has been successfully applied to video as well, for example in forensics. Wikipedia actually explains both pretty well:

http://en.m.wikipedia.org/wiki/Superresolution

Author
Time

I'm in the process of downscaling a clip from the Star Wars blu ray, such that we can compare the upscaling results objectively by calculating PSNR's for individual frames. I will post the original HD, the downscaled, and upscaled clips here.

Author
Time
 (Edited)

I will post a few clips of the super resolution v2 soon.

Author
Time

@dickalan when I click on your link it just goes to collider.com

Author
Time

Could you try a shot or two from the Technidisc? It has aliasing, like the GOUT, but there's no DVNR, so you wont have to fight as hard to extract detail.

This signature uses Markdown syntax, which makes it easy to add formatting like italics, bold, and lists:

Author
Time

Either the Technidisc or the JSC would work fine for this purpose.

Nobody sang The Bunny Song in years…

Author
Time

Of course, if you can send me a clip I'll run it through the algorithm.

Author
Time

Today I will also start upscaling the NTSC version of the GOUT. As You_too said, it apparently has more detail than the PAL version, and it's therefore better upscaling material. It will also provide for a fair comparison with Team Blu's impressive results.

Author
Time

DrDre said:

For those interested in finding out why multi-frame super resolution is possible, the first chapter of this phd thesis provides some explanations:

http://www.robots.ox.ac.uk/~vgg/publications/papers/pickup08.pdf

 As AntcuFaalb correctly pointed out, and as you can see from every example in that paper, you need to have plenty of aliasing for this to work. Fortunately, the GOUT has plenty of that, at least in one dimension! For most sources that are scanned well however, the best you can hope for with such methods is just noise reduction.

-G