How to get rig of blending ?

[Le Furet]

Hello

I already posted that on Neuron2 Forum, but I think the problem is difficult enough to justify several posts :)

I have more and more videos with lots of ghosting. If you want to see an exemple, see there (look at the arm moving, even deinterlaced you'll see blends on 3 or more successive frames) : http://fterraza.free.fr/test.vob

I'm working at making Unblend more accurate, but Unblend is still very basic. It can work in some "good" situation : typically if the blended frame is the result of the blending of the previous and the next, those being non-blended.

I'm currently looking for theoretical papers dealing with this problem (hoping that they exist, but I'm not sure Fansubbers problems could interest the guys of http://citeseer.nj.nec.com for example :D )

Does anybody know some of them ? And what is the "theoretical" word for "blending" and "unblending" ?

[scharfis_brain]

oh....

no way getting this crap unblended, exept of selecting the good frames by hand...

[mf]

ooOUCH! The pain... Low budget animation.. film handled with little care.. bad transfer.. contrast oversharpening.. *barf*

[Le Furet]

You understand now why I want to find first a theoretical solution, then maybe make a filter...

And that's no film : that's an animation serie created in the 70s, the dvd is an official taiwanese (yes :D ) one.

[MfA]

Wow, citseer doesnt have a single paper on inverse telecine and neither does ieeexplore. Not a very popular area of research academically speaking ... even the papers for the format conversion algorithms used by philips dont mention anything but simple pulldowns.

BTW why are yall talking about frames? Are blends always at the frame level? (Seems rather arbitrary, if they are going to pull stupid shit like blending they might as well do it for each field seperately.)

[Le Furet]

I have samples where blending is frame-based, but for this precise example, I don't know if it's frame or field based. I'm afraid it's frame-based (I must recognize that I don't really understand how they did this shit. I have seen such things (but not so bad) in bad NTSC->PAL tranferts, but the original was japanese and the dvd NTSC.

[mf]

Originally posted by Le Furet
And that's no film : that's an animation serie created in the 70s, the dvd is an official taiwanese (yes :D ) one.
You're not going to deny that it was originally recorded on film, are you?

[Le Furet]

I'm not good enough in english, and particulary forum english, to determine with precision what you call "film".

I just wanted to underline that I'm not sure the original clip was @ 24 fps.

I've just seen that the topic was moved. I'm not sure that's a good idea (despite all the respect I have for moderators). I didn't post to ask for an avs script, I think we all agree to say that no filter can help.

I posted to ask for an idea of the way (mathematic and algorithmic) to handle this sort of problem : how to detect on the frame what is ghost and what isn't, and how to remove it ?

[MfA]

The original frames/fields are always available somewhere in the stream?

If there are really 3 successive frames in which both fields are blends then that seems unlikely ... personally I think the only way to clean up something that badly fucked up would be to use a motion compensated median filter (which we unfortunately dont have).

[Le Furet]

The original frames/fields are always available somewhere in the stream?

I'm afraid that no...

personally I think the only way to clean up something that badly fucked up would be to use a motion compensated median filter (which we unfortunately dont have).

Even if this solution is of no use here, I am currently trying to develop a filter based on this idea (thanks to Manao).

But I need another approach... Maybe develop Ideas from Unblend (it tries to determinate the number t such that FrameN = t*FrameN-1 + (1-t)*FrameN+1 - The problems here are that more frames must be taken into account, and that we need to reconstruct the frame, whereas Unblend simply discard FrameN and replace it with FrameN-1 if t<0.5, and frameN+1 else).

[MfA]

IMO if you want to reverse blending where the original fields arent somewhere in the stream you should forget completely about trying to find an automatic algorithm without knowing the exact way in which the format conversion was done. Otherwise you are looking for magic ... if you cant even make a good guess about what the conversion process was dont expect any algorithm to do better.

You need to know the process to invert it (if it is invertible at all). So does anyone have a good guess at what exactly they did to produce those fields?

[Mug Funky]

OMFG! what were they THINKING??

there seems to be a couple of things going on here... i suspect the original telecine was done with "peculiarblend", that is a video camera pointed at a screen (i suspect almost all telecine was done this way before digital technology made it easier to go frame-by-frame), but there seems to be some kind of temporal smooth going on as well. like Johnny Cage's "shadow kick".

if the exact properties of this blend are known, it might be possible to reverse it, but you'd incur some serious rounding errors (not much of a problem). you'd still have peculiar blending to deal with, but that at least can be handled per field, rather than per 5-6 fields.

i'll see if i can't get something out of this.

[edit]
hmm. each frame appears to have a 1/6th blend of the previous.

i tried nested subtracts and got rid of most of it, but with intolerable side-artifacts (the picture was a LOT darker, and "pre-ghosting" appears on some frame-changes)

[zettai]

It's not eactly perculiar blend though as the amount of blending changes towards the middle.

I have a feeling that this type of telecine (which is used in a lot of anime, most noticably that filmed by Gainax pre-1999) is what is called polygonal prism telecine. There's a description of it here. (http://groups.google.co.uk/groups?dq=&hl=en&lr=&ie=UTF-8&selm=3ee46a3c.16771952%40news.reith.bbc.co.uk)

I've always thought that it must be possible to correct but finding out the pattern is very tricky.

It's seriously annoying. I've got hundreds of samples of it though, so plenty to use to develop a filter to overcome it :)

[Mug Funky]

interesting.

though the half-frame blend thing is produced by videotaping a projected image (i have some samples of this... obvious samples as there is a visible trapezoidal distortion from an off-centre camera, and before you ask, these were broadcast on australian TV, so they're not dirty pirate movies :))

i'll have to read up on this prism thing.