Why capture at 704x480 when all you want is 352x480 - noise!
 

I have been doing some reading on the net about resolutions for capturing VHS and TV. Almost everone says that you should capture at your final desired resolution. Almost everyone also suggests that the "proper" resolution for VHS is 352x480. This is what I've been doing for some time but I found something else interesting that caused me to do an experiment. Here we go:

The Basis of the theory:

Most capture cards actually capture each line at a very high resolution multiple times then blend those hi-res caps into one and then resize to your specified resoluion and pass them to the codec. There is no point (they say) in capturing VHS over about 480 because the VHS tape cannot hold more detail than that.

The Theory:

If you think about it, when you are capturing from a VCR, you are actually capturing two signals: 1. the vhs tape, 2. noise. Unlike the VHS tape, noise is not limited by resolution. Noise is entirely analog and thus has a very high resolution. So by requesting a frame width of 352, the capture card must blend the noise into the frame giving them both the same resolution. When it comes time to remove the noise, you will need to sacrifice video quality because in order to blend out noise, you need to blend out detail. By capturing well above the required resoluion, you can filter out the highfrequency noise without touching the underlying detail.

My test:

My test was done on a TV bradcast because the noise level on my "favorite" station is very high. Now I've been capturing from the station for some time with poor to OK results. In the last couple of days, I've been capturing at double my normal resolution (704x480). I apply my filters as I usually do and the results are amazing! After filtering, the cap looks very close to DVD. The noise is almost entirely eliminated yet the resolution is preserved. If I scale down to 352x480, well...

So what does everyone think? Am I nuts? Does anyone else do this?

Perhaps I could try a side by side comparison with the two different caps. I did my test on the Matrix comercial. I should be able to capture it again at my normal resolution and then compare the post filter results.

.
.

With VHS capturing, you also have to factor in, the resizing method, IF you
resize. There is nothing wrong w/ resizing down the 352 x 480. I've done
this many times. But, I let TMPG do it for me. ONLY, if the source size s
greater ie, 720 x 480. IMO, there seems to be a cleaner approach, and the
video looks more stirdy. Sorry, I can't explain this. Anyways.., yeah.
.
.
There are many way/approaches for resizing. And, that all depends on a
few factors, like

* the source
* capture card used
* editing method
* if cropping was done
* if letterboxing was done
* if ANY filteirng was done, as these will not always release the frame w/in
...a given boundry and result in crude frame reproduction etc. Yes, filters
...do effect this area too.
* resize in scripts or vdub or tmpg, and HOW source is fed into the final
...encoder for procesing.
* and, of course, the user behind the Start button :)

So, these are some more things to think about in VHS sources.

If you're up to it, why not use a better source VHS for your target testing.
Record a DVD movie onto a TAPE and test that ways. For now, just forget
about TV source on your Tapes. That is, if your isuses apply w/ respect to
your goal(s) :)

..me being theoriticle so more,
-vhelp

Re: Why capture at 704x480 when all you want is 352x480 - no
 

I'm trying to visualize your theory (assuming I understand correctly). The noise is composed mostly of high frequencies which get "aliased" when capturing at 352x480, but not at 704x480:

So there's a better chance of getting the original signal when capturing at high resolution, filtering, then resizing the image. Something similar to the way sound cards work when capturing audio, there's a low-pass filter that removes the noise before the signal is digitized (I suppose video cards could and probably do the same thing, but the filters we have available with AviSynth might do a better job).

You have it dead on. I am not saying that this is true but "assuming" it is and capturing and filtering accordingly, you may just end up with a near perfect capture.

Now I realize that my test was done on a TV capture but the same issue applies because the nature of the "resolution" of the noise. By capuring at 352x480 (like I have been) I cause the video card to blend the noise back into the frame thus making the size of the noise "wider" just as your diagrams show. With TV caps, this is easy to test:

1. Go to a noisy station and capture at 352x480.
2. Recap at 704x480.
3. Watch both at 2X in Vdub and compare the "size" of the noise.
4. Better yet apply the same Avisynth filtering script to both and compare the results.

This is just a theory. It could also be that the filters I am using are better working on 704x480 but either way, the result was much much better.

Let me know what you think? (Perhaps there's a Huffy mod codec in here.)

And thats exactly what I demonstrated refering to captures in my Thread "How funny" ...

http://www.kvcd.net/forum/viewtopic.php?t=6863

1. To scale down the captures to half horizontal size (I don't know "practically NOT theoretically" if its the same output if already caputuring at half horizontal depending of the internal card resizing algorhythms)
2. performing the avs filters ... and it really seems that they do perform better cause of less data to handle at half width
3. Even if re-sized back to 704x576 the quality is very very good in comparison to an output where the filters did perform on the untouched 704 width

@incredible

Actually I think we are talking about opposites. In fact, I think our ideas and results are in conflict. The key difference is that I am noise filtering in full resolution (your interlacing script and PixieDust) and you are resizing down before noise filtering. The only way that I can explain the opposite results would be to accuse the capture card's resizing circuit of not handling detail as well as Avisynth's routines.

I hadn't read your thread before so thanks for bringing it to my attention. Very interesting...I need to do more tests...

Re: Why capture at 704x480 when all you want is 352x480 - no
 

Similar to the sampling theorem, where you sample at least twice the highest frequency. Anything lower will cause distortion.
So it's better, at the expense of time, to capture at the original source resolution (if possible) and then do all filtering at the same captured resolution, where all the details + noise are "intact".
Then do any scaling you want.

-kwag

I think most of the work in incredible's version is done by the denoiser (PixieDust in the first script), the resizes from 704-352-704 basically perform a blur on the image. Otherwise there wouldn't have been as much of a difference when changing from PixieDust to VagueDenoiser or PeachSmoother. (I think Dialhot was saying something similar.)

Removing as much of the noise as possible before blending adjacent pixels together seems to make more sense to me, but as kwag pointed out, you do it at the expense of time (and space, when you capture at 704 instead of 352).

I have some material on VHS I want to capture, I'll have to do some tests using both theories sometime this week...

Yes the main job is done by pixiedust shure ... ;-)
But I showed there that it doesn't have the same effect if you perform the same denoising on the whole width. The meaning there is that the filter performs better at half width in comparison half width/filter full width/filter. Shure there's a touch of blur after this, but on the other hand I didn't plain the image as the image there was very noisy
http://www.kvcd.net/forum/viewtopic....9f92c32ef2cec6
(There directly the comparison starts)

BTW: ... the image wasn't interlaced

Recently I've done the down-upsampling from 696x576 -> standard bicubic 448x544 (filtering here with Deen only, using UnFilter(60,0) , 2 overscan blocks = 32 pixels off both height and width) -> standard bicubic 672x544 + 16-pixel borders on each side.

Looks very good, I was surprised that I didn't need a sharp bicubic or Lanczos to regain any sharpness in the final resize.

:) In a few words. Practise does show what does result (and that's also the intention of this thread I think) ... and in our purpose of KVCD ... what we will see on TV is what counts.
But every person gots his own "gusto" so ... everybody decides by himself
:)

At the risk of sounding obvious, I think it all depends on the source. Clearly, if we could apply Pixie to the original capture at full res and it looked better, we would do that. But, I think that downsizing and resizing itself does a lot for stability of the image. So, incredible is on to something here. (I noticed this effect when I accidently encoded at 352x240 capture at 352x480 but I hadn't done the testing and developement the incredible has)

Personally, all theories aside, I'm going to try both methods on each of my captures and pick the one that looks better.

I have also been testing capturing at 704x480 and resizing in Vdub down to 352x480 during capture. This can save a lot of hd space and since I'll be doing that anyway, why waste the space. I have to admit, the captures look better with the resizing than they do when I capture at 352x480.

And the nature of the noise, too. The noise caused by poor TV reception will be different than the noise from VHS over composite cable, which will be different than the noise from S-VHS over s-video cable, etc.

I usually fine tune the filters I use for each capture I do, so I'll probably end up doing this, too... 8-)

Risk? Absolutely not! :) Youre def. right, .... there never will be a "general" or optimal "for everything" script. The case there shown in the thread really begs for some heavy-anti-noise-treatment, but my "credo" is to kill noise by still keeping details or to get the best compromise (well thats what we all want to perform and receive, so nothing special).
As in the forum many scripts are offered for diff. purposes... the users will decide by themself which performs best seen on each source individually.
Well it blurs a bit but it has to be seen as a compromise between keeped details and on the other side noise/off. But I figured out that you have to perform the sharpen afterwards very individually .... as you already mentioned "everything depends on the source".
But the purpose of the script there was for very noisy sources ... and as 80% of my captures aren't broadcasted that noisy I wrote another script routine including Peachsmoother. As also posted there (page 3 down in the thread I think .. look for the "sat1" interlaced sample) and it performs faster and very well on full 704x576 less noisy res. without downscaling before. Boulder used a OneHalfOne on VHS sources but deen used instead of peachsmoother and IMO it came out very well too in case of VHS source. Well just try.