Perfect deinterlacing

[Megafox]

Prompt the Avisynth plugin for perfect deinterlacing or IVTC lossless. Half of the plugins that I found either duplicate the fields (terrible quality), or compensate for the movement (poorly visible doubling and degradation of quality), or perform mixing fields (ghost). But there are also plugins, for example Decomb (Has a poor search function that leaves some interlaced frames with any range and good spoils) and TIVTC (The search mechanism is better, but still, where the comb is very small - it remains, and 1/1000 good with bad quality ones I have not yet found plugins that could perfectly restore the video, I analyzed it in VirtualDub, and it shows that every 2 out of 5 frames are interlaced, and if there are less movements, only 1 of the comb is noticeable, or even unnoticeable. So why I did not find the plugin, with such a simple search engine? Or is there such a setting in existing ones? Translated through Google, do not be surprised

[Bernix]

Have you used BOB in deinterlace filters? Try ordinary yadif with temporal and spatial check. But if real 25i (50 fields) video in your case it is clearly NTSC, with bob you should get 50fps and all frames unique. So you have to calculate with higher bitrate. Also get right field orders, that is common mistake.

Bernix

[jagabo]

There's no such thing as perfect deinterlacing and there never will be. Right now QTGMC() is the best for most material -- by far.

2 out of 5 frames are interlaced

But your video should be inverse telecined, not deinterlaced: TFM().TDecimate(). Or, if a field blended NTSC/PAL frame rate conversion was used QTGMC().SRestore().

[hello_hello]

What source filter are you using to open the video with Avisynth?

You should post your script and upload a sample otherwise it's all guesswork as to what the problem is.

[johnmeyer]

Deinterlacing ALWAYS degrades the video. It is unavoidable. It cannot be otherwise.

My recommendation has always been this: do NOT deinterlace unless you absolutely have to. If your video is interlaced, save it as interlaced and then use the deinterlacer in your TV when you watch it. These are actually quite good these days and are optimized for the hardware on which the video is displayed.

Of course there are some situations where you MUST deinterlace, such as when you re-size interlaced video. In that case, you have absolutely no choice.

But, at the risk of repeating myself: if you don't absolutely need to deinterlace, don't do it because you will always lose quality. This can't be helped because you either have to discard information or you have to create new video by some sort of algorithm that estimates or averages or throws things away, etc.

[davexnet]

Deinterlacing ALWAYS degrades the video. It is unavoidable. It cannot be otherwise.

My recommendation has always been this: do NOT deinterlace unless you absolutely have to. If your video is interlaced, save it as interlaced and then use the deinterlacer in your TV when you watch it. These are actually quite good these days and are optimized for the hardware on which the video is displayed.

Of course there are some situations where you MUST deinterlace, such as when you re-size interlaced video. In that case, you have absolutely no choice.

But, at the risk of repeating myself: if you don't absolutely need to deinterlace, don't do it because you will always lose quality. This can't be helped because you either have to discard information or you have to create new video by some sort of algorithm that estimates or averages or throws things away, etc.

I think the intended use of the video is the key. For example, if you're going to upload a video to Youtube or some other video hosting site,
leaving it interlaced causes all kinds of problems, or at least it has in the past.

[manono]

If he's dealing with hard telecine he almost always should IVTC and not deinterlace. If some fine combing remains, then he can tighten the CThresh (in TIVTC) and/or make other adjustments so more of those are picked up. If, when those are post-processed (deinterlaced), then he can have TFM use a different and better post-processor.

But, to be sure, an untouched sample straight from the DVD should be provided - ten seconds of steady movement.

[hello_hello]

Deinterlacing ALWAYS degrades the video. It is unavoidable. It cannot be otherwise.

My recommendation has always been this: do NOT deinterlace unless you absolutely have to. If your video is interlaced, save it as interlaced and then use the deinterlacer in your TV when you watch it. These are actually quite good these days and are optimized for the hardware on which the video is displayed.

Of course there are some situations where you MUST deinterlace, such as when you re-size interlaced video. In that case, you have absolutely no choice.

But, at the risk of repeating myself: if you don't absolutely need to deinterlace, don't do it because you will always lose quality. This can't be helped because you either have to discard information or you have to create new video by some sort of algorithm that estimates or averages or throws things away, etc.

Double frame rate de-intelacing often allows for a substantial reduction in resolution without losing any (or very little) picture detail, based on the theory that 1080i and 720p have roughly the same picture resolution. I often resize PAL to 640x480 after de-interlacing and cropping, and when de-interlacing with QTGMC, mostly I'd say the encoded version looks better.

When encoding as interlaced, which if I remember correctly is something like 10% less efficient than progressive encoding (x264), it still has to be deinterlaced on playback, which I don't think manages to avoid any of the previously mentioned "problems".

In the case of QTGMC, I'd agree it changes the quality just as any other filter does, but more often than not I find it's de-interlacing to be an improvement so I don't mind "baking it in". It has a lossless mode where the original fields are retained. Even Yadif retains the original fields when using mode 1.

Anyway, if the OP only needs to remove the pulldown, I can't imagine why it wouldn't be better to do so.

[johnmeyer]

To be clear, when I said that deinterlacing ALWAYS degrades the video, I was NOT including the quality reduction from the re-encoding. Even if you save to uncompressed, or frameserve the result, it will still be degraded. The OP (and others) need to understand that there is no magic that can create something that wasn't there to begin with. Each method, whether bobbing, or motion estimation, or anything else, has to create pixels that weren't there before. Some (like QTGMC) are smarter than others, but even the smartest algorithm cannot do the job perfectly. There are ALWAYS artifacts, and sometimes they are quite easy to see.

Yes, the TV set will create some of these same artifacts when it plays the interlaced video, but I have argued for quite some time that the built-in hardware deinterlacers in most good TV sets are remarkably good. Put another way: I have seen bad deinterlacing in hundreds of posts in this forum and am guilty of producing bad results myself. By contrast, I have never been conscious of any artifacts while watching TV on my five-year-old Samsung 55" LCD, even though a lot of my content is 1080i.

[pandy]

Deinterlacing is lossless operation from video quality perspective. It may not meet expectation i.e. not able restore progressive video (before interlacing) quality.

[LMotlow]

Deinterlacing is lossless operation from video quality perspective.

Not true, and not possible. The laws of physics don't support that statement. The second part of the above post is obvious.

[pandy]

Deinterlacing is lossless operation from video quality perspective.

Not true, and not possible. The laws of physics don't support that statement. The second part of the above post is obvious.

It is true and inline with math and physics - your claim is like blaming interpolator for not being able to extend signal bandwidth... (hope you understand why resampling for example 22kHz signal to 44kHz will not extend signal itself bandwidth). Deinterlacing is special form of signal resampling (interlacing is founded on principle where every second sample is removed and this violate Nyquist law) and perfect deinterlacer doesn't exist as quest for perfect reconstruction of signal before sampling is more like artistic educated guess than linear math.

[LMotlow]

Deinterlacing is lossless operation from video quality perspective.

Not true, and not possible. The laws of physics don't support that statement. The second part of the above post is obvious.

It is true and inline with math and physics - your claim is like blaming interpolator for not being able to extend signal bandwidth... (hope you understand why resampling for example 22kHz signal to 44kHz will not extend signal itself bandwidth). Deinterlacing is special form of signal resampling (interlacing is founded on principle where every second sample is removed and this violate Nyquist law) and perfect deinterlacer doesn't exist as quest for perfect reconstruction of signal before sampling is more like artistic educated guess than linear math.

That's a new one.

[pandy]

That's a new one.

So... seem you blaming resampler for not extending signal bandwidth after upsampling - why? Are you able to elaborate on this?

[johnmeyer]

Deinterlacing is lossless operation from video quality perspective.

Most definitely not true.

The two fields in interlaced video come from different moments in time and from different points in space. To create progressive video, you must either create a field that occupies a different point in space, or a different moment in time. Since there is no way to know, perfectly, what and when that field should be, the result will ALWAYS have artifacts no matter how good the algorithm.

I guess we can get into a discussion about what "lossless" means, but to me it means that if pixels are changed, the result is not lossless. I think that some in this thread (and elsewhere) equate the words lossy or lossless only with the encoding operation. However, as I posted above, even if you encode using uncompressed or encode using a lossless codec, your video will still always be degraded by the deinterlacing process.

Another way to get at this is to look at the thousands of posts about deinterlacing and note that a large number of those posts point to QTGMC as being one of the best deinterlacers anyone has used. The reason that it is so good is that it does both adaptive deinterlacing and also uses motion estimation to construct the needed fields. But QTGMC obviously uses motion estimation, which is notorious for the artifacts it creates, and which constructs fields by estimating where pixels should be placed.

So, video quality is always affected when deinterlacing.

[Bernix]

Understand properly, that movie recorded as progressive (or simpler on classical film) then translated as interlaced, has each field from different time ?
Have to think twice next time
Thank you

Bernix

[johnmeyer]

Understand properly, that movie recorded as progressive (or simpler on classical film) then translated as interlaced, has each field from different time ?
Have to think twice next time

Not quite. A 24 fps movie is translated to 30 fps video by adding pulldown fields. These are simply duplicates of existing fields. If you want to recover the original 24 fps movie, you can do that without any loss or artifacts by doing inverse telecine, sometimes abbreviated as IVTC.

It is very common for people who are first confronted with interlacing to confuse deinterlacing with inverse telecine. However, they are completely and totally different. The only similarity is that both operations yield video that is progressive. Inverse telecine simply removes duplicate fields that were added so that the 24 fps film can play on devices that are hard-wired to play at 30 fps; deinterlacing creates new fields that didn't exist before. So, one removes fields and the other creates fields.

And yes, 24 fps is usually 23.976 fps and 30 fps is usually 29.97 fps, thanks to the 1000/1001 multiplier used by NTSC video.

[manono]

Bernix says he's in Europe and, I'll grant you, his English is tortured. But I read his post as having nothing to do with telecining but perhaps more to do with phase-shifting, a not-uncommon phenomenon in Europe, especially with television captures as well as DVDs. In those cases the films look, at first glance, to really be interlaced.

I think he's saying that for films, each field won't be from a different time, and therefore the film isn't really interlaced. In such cases a simple field matcher (such as TFM()) will serve to make it progressive again without the damage a deinterlacer will do. He might also be referring to field-blended garbage where deinterlacing isn't the way to go, but you'll need an unblender instead.

But these are just guesses as the post isn't all that clear (to me).

[Bernix]

Hi Manono,
you are right, my english is puzzle... I never had english at school and my language is based on "words elasticity" we have 7 declension and much more. So definitely different from english language. And add my dumbness and you have cypher that german enigma never had

Bernix

[hello_hello]

I guess we can get into a discussion about what "lossless" means, but to me it means that if pixels are changed, the result is not lossless. I think that some in this thread (and elsewhere) equate the words lossy or lossless only with the encoding operation. However, as I posted above, even if you encode using uncompressed or encode using a lossless codec, your video will still always be degraded by the deinterlacing process.

If you bob deinterlace with a deinterlacer that doesn't touch the original fields, Yadif(mode=1) or QTGMC(lossless=1) etc, and you encode losslessly, surely it must be possible to extract the original fields at a later point in time. Not that I'm sure I'd ever want to do so, but in theory you should be able to.

Another way to get at this is to look at the thousands of posts about deinterlacing and note that a large number of those posts point to QTGMC as being one of the best deinterlacers anyone has used. The reason that it is so good is that it does both adaptive deinterlacing and also uses motion estimation to construct the needed fields. But QTGMC obviously uses motion estimation, which is notorious for the artifacts it creates, and which constructs fields by estimating where pixels should be placed.

Doesn't your TV do much the same thing when de-interlacing?

Given I almost never watch video any other way than decoding it on my PC for displaying on the TV, I thought I should try it for myself, so I remuxed a vob file as a TS file (I'd already ripped it from a PAL DVD for encoding), stuck it on a USB stick and gave it to the TV's media player. The TVs de-interlacing didn't blow me away. Probably about the same quality as Yadif in double frame rate mode, which is how I remembered it. Not horrible, but not as good as QTGMC, in my opinion.
I don't know if the TV would do a better job if I fed it an interlaced PAL or NTSC signal via the Bluray player, or maybe the Bluray player might de-interlace better, but that's not how I watch video anyway. The TV is a Samsung Plasma. About 6 years old.

[poisondeathray]

Given I almost never watch video any other way than decoding it on my PC for displaying on the TV, I thought I should try it for myself, so I remuxed a vob file as a TS file (I'd already ripped it from a PAL DVD for encoding), stuck it on a USB stick and gave it to the TV's media player. The TVs de-interlacing didn't blow me away. Probably about the same quality as Yadif in double frame rate mode, which is how I remembered it. Not horrible, but not as good as QTGMC, in my opinion.
I don't know if the TV would do a better job if I fed it an interlaced PAL or NTSC signal via the Bluray player, or maybe the Bluray player might de-interlace better, but that's not how I watch video anyway. The TV is a Samsung Plasma. About 6 years old.

That's "normal". Because >99% of HDTV's use a simple bob deinterlace. That's it. So it looks like a bob or yadif.

The more expensive sets use additional processing, and they will usually advertise it under something like "motion adaptive deinterlacing" or some fabulous trade name for "special deinterlacing". Extra processing features are things companies can charge for

Bob or yadif is good enough for most people. But put a cheap TV side by side to a more expensive set , or QTGMC processed video, and even "joe public" can see the difference. Another quickie way is to pause the picture . Cheap sets will show the aliasing. Expensive sets will attempt to fill in the gaps, antialias or other processing

[johnmeyer]

You can tell if your TV has one of the better deinterlacers if it also offers the "soap opera" effect where they can make movies look like TV video. Almost everyone turns that off, because it looks so awful, but the same technology can be used to provide "QTGMC-like" deinterlacing.

[jagabo]

If you bob deinterlace with a deinterlacer that doesn't touch the original fields, Yadif(mode=1) or QTGMC(lossless=1) etc, and you encode losslessly, surely it must be possible to extract the original fields at a later point in time.

Yes, it is possible. For example:

WnateverSource("filename.ext") # get a YV12 video
#AssumeTFF() or AssumeBFF() if necessary
src=last # remember it

QTGMC(lossless=1) # double frame rate deinterlace, keeping orignal fields, or Yadif(mode=1)

SeparateFields() # weave the original fields back together
SelectEvery(4,0,3)
Weave()

Subtract(src,last).Levels(120,1,136,0,255) # show the differences, accentuated

You will see a plain, flat grey video (indicating no differences). Remove the "lossless=1" and you will see many differences (as deviations from grey).

[manono]

Hi Manono,
you are right, my english is puzzle

A puzzle, indeed. But your English is way better than my German and I was 100% fluent in German when still a child. And I'm sure your German is easily as good as my English. So, no problems.

[pandy]

Deinterlacing is lossless operation from video quality perspective.

Most definitely not true.

So, video quality is always affected when deinterlacing.

Once again - please explain why you expecting from upsampler, original (before sampling), signal bandwidth reconstruction.

Please use Whittaker–Nyquist–Kotelnikov–Shannon sampling theorem as foundation for this discussion.

Deinterlacer is special case for reconstruction filter in presence of (usually) undersampled signal (usually as i can imagine special preprocessing before interlacing to join motion and bandwidth to satisfy Whittaker–Nyquist–Kotelnikov–Shannon sampling theorem however for most of situations signal before interlacing is not properly bandwidth limited i.e. it will violate in unavoidable way Whittaker–Nyquist–Kotelnikov–Shannon sampling theorem).

So once again why you expect from interpolator (upsampler) reconstruction of the original signal bandwidth (before undersampling)?

My claim is: Any sane deinterlacer (not changing original signal samples) is lossless from signal perspective even if deinterlacer deliver ugly (subjectively) video. Use as reference progressive video and use as lowest common denominator (no deinterlacer) video with missing half of lines (odd or even) - fill gap between lines with some solid colour - do some SSIM, VMAF, PSNR or whatever metric you like tests.

I can imagine upsampler that based on provided information (signal samples) may more or less accurately resynthesize some imaginary signal spectrum but this belongs purely to subjective area not Whittaker–Nyquist–Kotelnikov–Shannon sampling theorem.

I know your claim but IMHO there is different (than

Deinterlacing ALWAYS degrades the video. It is unavoidable.

) tree to bark - don't blame deinterlacer as you will be contradictory to yourself (

It cannot be otherwise.

) - deinterlacer can't be blamed for something beyond his responsibility.

[johnmeyer]

Once again - please explain why you expecting from upsampler, original (before sampling), signal bandwidth reconstruction.
<snip>

My claim is: Any sane deinterlacer (not changing original signal samples) is lossless from signal perspective even if deinterlacer deliver ugly (subjectively) video. Use as reference progressive video and use as lowest common denominator (no deinterlacer) video with missing half of lines (odd or even) - fill gap between lines with some solid colour - do some SSIM, VMAF, PSNR or whatever metric you like tests.

None of this makes any sense at all. I have no idea what you are trying to say. My only point is that deinterlacing introduces artifacts. It is a cold, hard fact. I really have tried to understand the point you keep trying to make, but your posts are extremely confusing and difficult to understand.

Why are your posts confusing? Because you totally agree with the only point I have been making, but despite that agreement you keep posting about something that has nothing whatsoever to do with anything I have posted. It appears that the only reason for your posts is that you are very hung up on the word "lossless."

To prove that we are saying the same thing, I look at what you just posted:

"Any sane deinterlacer (not changing original signal samples) is lossless from signal perspective even if deinterlacer deliver ugly (subjectively) video."

Thus, we both agree that the deinterlacer can create video with artifacts. That is the only point I have been trying to make. I have made no point about lossless, other than when I tried to differentiate between changes made during encoding, and changes made by AVISynth filters.

So, it is clear that you understand that the deinterlacer can deliver "ugly" video. Since we agree on that, and since that is the only point I have been trying to make, I see no reason to continue posting, at cross purposes, over the definition of the word "lossless." It is a totally pointless discussion, and serves no purpose whatsever to continue it. I am certainly not going to post any more in this thread.

[pandy]

My only point is that deinterlacing introduces artifacts. It is a cold, hard fact. I really have tried to understand the point you keep trying to make, but your posts are extremely confusing and difficult to understand.

Well you claimed that deinterlacer work are not lossless operation from signal perspective - which is wrong assumption from my point - we can't ignore fact that deinterlacer is a solution to fill a gap where signal not exist - you have right to be confused - my point was strictly that you should not blame deinterlacer for being unable to recover non existing information that was brutally removed from signal. You may complain on deinterlacer insufficient "imagination" or "creativity" to fill those gaps however deinterlacer not changing signal at all.

Why are your posts confusing? Because you totally agree with the only point I have been making, but despite that agreement you keep posting about something that has nothing whatsoever to do with anything I have posted. It appears that the only reason for your posts is that you are very hung up on the word "lossless."

To prove that we are saying the same thing, I look at what you just posted:

"Any sane deinterlacer (not changing original signal samples) is lossless from signal perspective even if deinterlacer deliver ugly (subjectively) video."

Thus, we both agree that the deinterlacer can create video with artifacts. That is the only point I have been trying to make. I have made no point about lossless, other than when I tried to differentiate between changes made during encoding, and changes made by AVISynth filters.

So, it is clear that you understand that the deinterlacer can deliver "ugly" video. Since we agree on that, and since that is the only point I have been trying to make, I see no reason to continue posting, at cross purposes, over the definition of the word "lossless." It is a totally pointless discussion, and serves no purpose whatsever to continue it. I am certainly not going to post any more in this thread.

Pointless is to make false (IMHO) statement and later rejecting to support or deny those statement. Perhaps it is difficult to understand me due my English (being not native in English and it is purely self learned ) - being engineer i see problem in slightly different light - deinterlacer can't be blamed for destroying signal quality - signal quality was destroyed when half of lines was removed and you can't blame for those line removal deinterlacer.
From signal processing perspective however every deinterlacer not affecting existing signal samples (lines) is lossless operation even if video is ugly (like doubled lines or missing lines) - it is lossless as original signal is not affected - you simply don't like artistic effect created by deinterlacer.
And IMHO word "lossless" is very important on site like this especially when we talking about signal processing.

So once again - deinterlacer is lossless operation from signal perspective.

[Bernix]

Hi,
I'm also not english speaker but understand I think well. Simply, the delivered data can be recovered because untouched. There is only add missing lines that can be removed again. When compressed lossless. Other way it is destructed by video codec not by deinterlacing filter. If I understand it right it is clearly explained.

Bernix

[Bernix]

Just for curiosity, has anybody tried to do it inverse? Remove transmitted lines and left in video just "created" lines? Of course field order will be opposite, just curious how such video will look.

Bernix

[pandy]

This is easy to perform experiment - take progressive source, interlace it, new progressive once again interlace (re interlace) however with different (opposite) interlace scheme. Depends on quality (type) of deinterlacer (in first deinterlace step) results may be different (from line doubled deinterlacer mean no loss on signal at a cos of ugliness of deinterlaced in first step video to pretty OK for most cases but never perfect - at some case it can be however constant field of solid color - case for replacing missing samples by some fixed value - usually in DSP a '0' (zero) is used to separate samples - black lines - selecting those lines we end with black video - all samples loss).

Personally believe that with sufficiently trained neural network close to perfect results can be achieved - but this is not trivial task - perfect results can be gained if machines will begin to perceive video as humans (so not signal/samples but object context) - if this is realistic? - have no clue - not my area.

And once again deinterlacing is a lossless operation so long as original samples (original field lines) stays untouched (i'm not aware of existence any deinterlacer that affecting/alter present signal lines to recover missing ones).