dvd to HD

[Sagittaire]

It's not possible to compare 576p with 1080p. If you're viewing the two videos under the same conditions (i.e. same viewable area), then the 576p was upscaled too. So you're just comparing one scaling algorithm with another. Maybe one of the algorithms is in software while the other is built into your video card or even the physical properties of your crt, but either way it's scaling.


Well compare 576p and 1080p with same area is possible (video projector for example) and in this case for same area 1080p will be visualy better than upscaled 576p@1080p and upscaled 576p@1080p will be better than 576p.


So the right question to ask is: which scaling algorithm looks best, and should it be applied before or after compression. Not whether to scale.


In fact the initial question is what is the best way ...

1) 576p DVD -> 1080p hardware upscaling
2) 576p DVD -> 1080p software upscaling -> 1080p reencoding

[shon3i]

1) 576p DVD -> 1080p hardware upscaling
2) 576p DVD -> 1080p software upscaling -> 1080p reencodingWhat you use in you samples here

Which script you use in encoding, because this results to me looks good. Like you say 576p@1080p looks better than 576p

[foxyshadis]

Well compare 576p and 1080p with same area is possible (video projector for example) and in this case for same area 1080p will be visualy better than upscaled 576p@1080p and upscaled 576p@1080p will be better than 576p.

That's just another form of scaling! Being analog, it can't really be classified as nearest neighbor or bilinear, but its effects fall somewhere in between them. Sort of similar to a hermite scaler onto a very high-resolution field. Any way you put it, comparing two things of the same size that come from different sizes, compares the processing to get them there (even if the sum total of the processing is lens distortion).

[fenomeno83]

????????????
Now I have confusion in my head!:D

Sagittaire please post your steps to get a HD movie from Dvd movie

[SealTooGreat]

What about blind deconvolution for improving image details?!
576p->1080p->blind deconvolution->final video
or
576p->blind deconvolution->upscale->1080p
I've seen pretty amazing result what blind deconvolution can do to poor_detail/blurry picture: http://math.nist.gov/mcsd/highlights/blinddeconvolution.html

[foxyshadis]

You can't really use deconvolution on compressed images/video to maximum effect, they really require raw, uncompressed source. If you try, you end up with DCT artifacts overpowering the content. However, if you really want to try it, download SAR (http://www.general-cathexis.com/) and toss a few screencaps at it. Otherwise, you can try limitedsharpen and seesaw, plus some combination of eedi2 and lanczos. (Like Multiswar and my EKG, and the original prototype in IIP.) LimitedSharpen is actually fairly similar to deconvolution, as you can see if you turn the strength all the way up.

[chadamir]

People do this all of the time with their htpcs using ffdshow with a dvd. I've never seen anyone want to do it with an encode though. The reason people bother is that their tvs are going to do a shitty job if they don't upscale it no matter what or their computer can do a better one. It's more of an effort to preserve detail than the hope of gaining any.

[reepa]

Upscaling can be useful in many circumstances. For example, if you have standard definition video, it will look better upscaled on a high-definition display than 1:1 pixel-mapped on a standard definition display, because on a HD display the physical pixels are smaller.

The same kind of goes for audio. You can convert 44.1KHz audio directly to analog and filter it before outputting it to the speakers, but you require a very steep filter to remove all the frequencies above 22.05KHz and pass all below that. Or, you can digitally resample ("upscale") to 88.2KHz (double the original sampling rate because integer resampling is easy), convert to analog, and use a simpler analog filter.

However, postprocessing video and then re-encoding it doesn't sound very useful since the disadvantages (compression artifacts, time) outweigh the advantages (better scaling algorithm).

Also, deconvolution is something totally awesome. Especially the blind variety.

PS: it should be obvious that upscaled doesn't match something that originally has a better resolution (upscaled SD vs HD for example).

PS2: this isn't based on mathematical knowledge, but when resampling, it should be done by a large amount. For example, 1280x720 to 1366x768 is bad (damn those LCD televisions!), 720x576 to 1920x1080 is less bad. Another example is older Creative sound cards: they resample everything to 48KHz (most music is 44.1KHz), which is considered bad.

[PaulKroll]

Even assuming a series of frames could be used to extract out some details, given the right algorithm, you still have the problem with the source. Any MPEG (1/2/4/name it) stream gets part of its compression by not duplicating areas of the frame that are the same in a previous or later frame. You'd end up with an image that had large, blocky sections that were still blurry, relative to the suddenly detailed sections.

Use something other than bilinear to upscale, sure, or a sharpener or a script like IIP meant to do a whole lot of processing on a piece of video. That can look considerably better than the original DVD, depending on your taste and how much time you'll put into it. (But it can look worse, too.)

Incidentally I downloaded and ran the "Video Enhancer" version of VirtualDub, rendering out a video using its "Super Resolution" to quadruple a video's res to 1440/960. Comparing it with an avisynth'd Lanczos4Resize'd version, it looks pretty much like they're identical. Except VE is drinking CPU for No Good Reason (it took a LONG time to render).

[foxyshadis]

PS2: this isn't based on mathematical knowledge, but when resampling, it should be done by a large amount. For example, 1280x720 to 1366x768 is bad (damn those LCD televisions!), 720x576 to 1920x1080 is less bad. Another example is older Creative sound cards: they resample everything to 48KHz (most music is 44.1KHz), which is considered bad.

All modern soundcards resample to 48/24 (or 96/24 at the pro end) before the A/D, including Creative's current X-fi. The real reason it was considered bad was that Creative's implementation was shamefully poor, all the way from AWE32 to Audigy, and it wasn't until X-fi that they got their act together. After all, we've had SSRC doing near-perfect conversions between 44<->48 for movies for years now, and M-Audio and other audio cards the same in hardware.

Video and audio sampling isn't really comprable, because of differences in band-limiting and filtering, but how much you resample doesn't really make a difference, if done properly. If you were to upscale 10% and then downscale back to the original, you'd be hard pressed to spot the difference, although each of the algorithms have their peculiar artifacts that get more obvious the more you expand/contract. The main reason to stick to small sizes is to keep from enlarging artifacts (existing and algorithmic) too much.

I personally think modified blind deconvolution is a neat trick that, like x264's esa option, other much faster methods should strive toward, rather than seeing it as a real solution.

[Morte66]

PS2: this isn't based on mathematical knowledge, but when resampling, it should be done by a large amount. For example, 1280x720 to 1366x768 is bad (damn those LCD televisions!), 720x576 to 1920x1080 is less bad.

Agreed. Resizing 720x576 to 1280xNNN on my old monitor, I had to use Lanczos to get acceptably low artefacts. Resizing to 1920xNNN, I'm OK with Bicubic. It seems that the bigger the target, the thinner the artefacts get spread. Eventually they look like fairly innocuous random noise instead of fairly annoying structured artefacts.

[reepa]

but how much you resample doesn't really make a difference, if done properly. If you were to upscale 10% and then downscale back to the original, you'd be hard pressed to spot the difference, although each of the algorithms have their peculiar artifacts that get more obvious the more you expand/contract.

I still think the contrary. Here's a simple experiment: create an image with a small checkerboard pattern (every other pixel black, every other white). It basically has the highest possible detail a given image can have. Resize it with your favorite algorithm by a small amount (upscale by 120% for example) and a larger amount (upscale by 170% for example). You'll probably notice that the image that was resized by a small amount has beat patterns, while the other image has none.

Larger upsampling should also help with audio. The larger the upsampling, the wider the passband of the analog filter can be, meaning less unwanted phase shift, emphasis, and unwanted harmonics. Upsampling from 44.1KHz to 48KHz allows for a 3.9KHz passband, while upsampling from 44.1KHz to 88.2KHz allows for a 44.1KHz passband. Here's how I understand digital to analog conversion (please correct me if I'm horribly wrong):

http://www.freeimagehosting.net/uploads/dafcb1e156.png (http://www.freeimagehosting.net/)

[*.mp4 guy]

The reason that happens to the image isn't because of the resizer, its because you are taking a square peg and trying to fit it in a round hole, as it were -> the problem isn't that the resizer messed up the image, its that your original image can't be stored at the destination resolution without artifacts, your image isn't representative of normal content, the highest possible quality resize you can get with it is with point resizing to double the resolution, a technique that looks awful on any normal image.

[reepa]

The reason that happens to the image isn't because of the resizer, its because you are taking a square peg and trying to fit it in a round hole, as it were -> the problem isn't that the resizer messed up the image, its that your original image can't be stored at the destination resolution without artifacts

I think that proves my point about the amount of resizing having an effect (a negative effect if difference between the source and the target sizes is too small).


your image isn't representative of normal content

Obviously, but I'm showcasing the worst possible scenario. A bit like when testing speakers you usually do a frequency sweep and measure it even though few songs actually contain a frequency sweep.

[*.mp4 guy]

I think that proves my point about the amount of resizing having an effect (a negative effect if difference between the source and the target sizes is too small).

I don't think you completely understood what I was saying, unless you are exactly doubling the resolution you get distortion since your input image is incapable of being stored losslessly at your destination resolution, however this should never be an issue with normal content as long as it doesn't have any aliasing in it, if there is aliasing, then there will be some noticible loss when resizing to anything but a multiple of 2 of the original resolution. Basically you get distortion during resampling if the original image had more spatial resolution crammed into it then it could actually hold (IE nearest neighbor downsizing) your image has the absolute maximum spatial detail that can be represented in an image of its size, no advanced filtering will know what to do with it (try compressing it with jpeg to get a good example of this, jpeg works great on normal images, but it will have a fit with yours and to a lesser extent any image with aliasing)

[akupenguin]

Actually, your example compresses quite well in jpeg. Checkerboard is one of DCT's basis-vectors, so there's only one nonzero coefficient per block, and no particular artifacts. :)

[*.mp4 guy]

If you say so...

dots.bmp(103KB) (http://img242.imageshack.us/img242/9195/dotsky2.png)
dots.jpg(103KB) (http://img234.imageshack.us/img234/4701/dots2hn6.jpg)

I have to say I never realised that bitmaps had superior compression and image quality compared to jpg.... When a bitmap can outperform jpeg in perceptual image quality, while being the same size I would say its a fair assumption that it performs poorly in this context.

[edit] imageshack recompressed the bmp image to a png, so you will just have to take my word that a monochrome bitmap of that image is 103KB.

I made a mistake in my last post, its not every multiple of 2 of the original resolution that is lossless, its any integer multiple of the original image.

[foxyshadis]

That's not a checkerboard pattern:

png (250b) (http://foxyshadis.slightlydark.com/random/checker.png)
jpg (15k) (http://foxyshadis.slightlydark.com/random/checker.jpg)
j2k (600b) (http://foxyshadis.slightlydark.com/random/checker.jp2)

The funny part is that the j2k compressor is smart enough to find and use the checkerboard pattern as it should, whereas most jpeg implementations are too dumb to check for the optimal pattern. They all use essentially the same quant matrix, as well; jpeg encoders are pretty rudimentarily developed overall.

[*.mp4 guy]

That's not a checkerboard pattern:

png (250b) (http://foxyshadis.slightlydark.com/random/checker.png)
jpg (15k) (http://foxyshadis.slightlydark.com/random/checker.jpg)
j2k (600b) (http://foxyshadis.slightlydark.com/random/checker.jp2)
The funny part is that the j2k compressor is smart enough to find and use the checkerboard pattern as it should, whereas most jpeg implementations are too dumb to check for the optimal pattern. They all use essentially the same quant matrix, as well; jpeg encoders are pretty rudimentarily developed overall.

I realise that (now), I used that pattern because it exhibits the same effects when being resized as reepa's example and I already had made it (to test monitor resolution on crts), I had forgotten that it wasn't a strict checkerboard pattern, however my point about jpeg doing badly on images of that type, is, I beleive still valid. akupenguin is of course right in theory, though even your ideal checkerboard pattern seems to have given the jpeg encoder a lot of trouble. I should have double checked my pattern when I got such bad results from jpeg; though the results on the true checkerboard pattern backup my point to a lesser degree.

I think the problem is the non-finite accuracy of jpeg dct->idct algorithms it causes errors at higher quants that should theoretically be lossless if perfect accuracy was achieved, H264's "quasi dct" transform would probably do much better since it has finite accuracy, in my experience H264 does do much better on patterns of this nature then previous mpeg standards.

[edit]
actually the only checkerboard pattern that jpg will handle correctly isn't an exact checkerboard, it fades at the edges, only the center has maximum contrast, I'm not sure about h264, I don't have a visual reference for its individual coeficients. visual representation of individual dct coeficients (http://img216.imageshack.us/img216/231/dctcoeficientsct5.png).

Also, foxyshadis your checkerboard has incorrect polarity, the first pixel should be white not black.

[foxyshadis]

Doesn't actually matter. A maximum 64th coefficient will render either correctly based on whether the DC coefficient is black or white. And with the 64th at maximum it should saturate enough that there's no fading at the edges. If the DC was grey you might be able get away with lower strength on correct polarity, but after huffman it'd be the same size anyway. (I don't know the exact point that sample images are at, probably around half, otherwise all jpeg blocks would fade to grey at block edges.)

Anyway, sorry for turning this away from resizing. As mp4 guy says, there's a huge difference between resizing from aliased images and more normal antialiased video/images - like some audio tends to jam it up to the nyquist limit and requires special handling, whereas in songs with a 16khz lowpass there's not as much difference between various resample methods. (Until you create junk up there with post-processing.)

A much better test of image resamplers is text overlay, which is usually the first thing to go fuzzy if the resizer isn't up to the task. It's very rare to see a resizer cause beat frequencies (unless you do a lot of print scanning), but very common to degrade text and line art.

[*.mp4 guy]

@ foxyshadis: doh, I should have realised that about the dc coeficient =/

To get back on topic, here is a preliminary test with multiswar v2, there are still some bugs to be worked out (the blinking line on the left, which is related to masking, and a color shift caused by one of the sharpening functions).

Heres the script, Ill list the dependancies and parameters when its finnished.
# MultiSWAR - Multi Step Warping And Resizing - V2 alpha1
#
#

Function MultiSWAR(Clip Clp,Int"DestX",Int"DestY",Int"StepRatio",Int"Sharpen", float "ssm", int "dblur", float "thr", float "dthr", int "smode", int "esharp", float "ss", int "sthresh", float "sstrength", int "sthresh2", float "sblur")

{
OX = m(16,Clp.Width)
OY = m(16,Clp.Height)

DestX = Default(DestX, OX*2)
Desty = Default(Desty, OY*2)
Stepratio = Default(Stepratio, 3)
Sharpen = Default(Sharpen, 20)
SSm = Default(SSm, 2)
dblur = Default(Dblur, 5)
thr = Default(thr, 1)
dthr = Default(dthr, 0.1)
smode = Default(smode, 4)
esharp = Default(esharp, 0)
ss = Default(ss, 1)
sthresh = Default(sthresh, 1)
sstrength = Default(sstrength, 0.8)
sthresh2 = Default(sthresh2, 255)
sblur = Default(sblur, 1)


Clp.Spline36Resize(OX+round(OX/StepRatio/4)*4,OY+round(OY/StepRatio/4)*4)
OXo = OX+round(OX/StepRatio/4)*4
OYo = OY+round(OY/StepRatio/4)*4
o = last.converttoyv12()
s = o.aWarpSharp(Depth=4, Blurlevel=1, cm=0)
ss = s.spline36resize(OX*SSm, OY*SSm)
Nodetail = ss.tbilateral(diameterl=dblur, diameterc=3, sdevl=thr, sdevc=thr)
mask1 = subtract(ss, NoDetail)
mask2 = mask1.levels(55, 1, 127, 0, 255, false).levels(127, 1, 255, 0, 127, false).levels(0, dthr, 127, 0, 255, false)#.blur(1)
mask3 = mask1.invert.levels(55, 1, 127, 0, 255, false).levels(127, 1, 255, 0, 127, false).levels(0, dthr, 127, 0, 255, false)#.blur(1)
mask4 = maskedmerge(mask3, mask2, mask3).spline36resize(OXo, OYo)#.undot.degrainmedian(mode=3)
lines = s.LimitedSharpenfaster(Smode=Smode, soft=8, strength=sharpen, overshoot=Esharp, SS_X=1.25, SS_Y=1.25, lmode=1).DCTFun4b(Sthresh,Sthresh).FrFun7(Sstrength,Sthresh2,Sthresh2)
detail1 = o.gaussianblur(vary=sblur, varc=sblur)
detail2 = subtract(o, detail1).invert
detail3 = subtract(o, detail2)
maskedmerge(detail3, lines, mask4)

return (width < destx) ? MultiSWAR(DestX, DestY, StepRatio=StepRatio, Sharpen=sharpen, ssm=ssm, dblur=dblur, thr=thr, dthr=dthr, smode=smode, esharp=esharp, ss=1, sthresh=sthresh, sstrength=sstrength, sthresh2=sthresh2, sblur=sblur) : spline36resize(destx,desty)

}
function m(int r, float x) {return(x<16?16:int(round(x/float(r))*r))}
and heres the script used for the example
Setmemorymax(768)
import("D:\Media\Plugins\my functions\MultiSWAR_V2.avs")

MPEG2Source("D:\rips\trailer\Trailer.d2v", cpu=0, idct=4).crop(0, 60, 0, -64)#.spline36resize(720, 304)

MultiSWAR(stepratio=4, sharpen=20, thr=0.5)#thr might not match what I used, I went through a lot of testing

#I recomend splitting the script in half at this point, run the first half, then feed its output into the second half (this provides a large speedup since otherwise the script becomes ram limited)

source = last.ttempsmooth()

backward_vec2 = source.MVAnalyse(isb = true, delta = 2, pel = 2, overlap=4, sharp=2, idx = 1, truemotion=true)
backward_vec1 = source.MVAnalyse(isb = true, delta = 1, pel = 2, overlap=4, sharp=2, idx = 1, truemotion=true)
forward_vec1 = source.MVAnalyse(isb = false, delta = 1, pel = 2, overlap=4, sharp=2, idx = 1, truemotion=true)
forward_vec2 = source.MVAnalyse(isb = false, delta = 2, pel = 2, overlap=4, sharp=2, idx = 1, truemotion=true)

source3 = source.MVDegrain2(backward_vec1,forward_vec1,backward_vec2,forward_vec2,thSAD=400,idx=1)
source3.gradfun2db(1.5)

Heres the result (http://www.megaupload.com/?d=9EKNF3HR), of the script, I'm presently uploading the source, I'll edit a link in once its done.
[edit]heres the source link (http://www.megaupload.com/?d=QS2DV1V8)
[edit2] Looks like yousendit requres registration to download now, I have uploaded the file to a different site and updated the links

[Chainmax]

Sharktooth: could you include three resizes: one made with EEDI2, Spline36 and LimitedSharpenFaster, one made with foxyshadis's EKG and one made with *.mp4 guy's MultiSWAR?

scharfis_brain: aah, the Gabriel Knight "Making Of" video :). I remember Didée saying it could be sped up a lot, too bad this awesome function was never developed further :(.

[Sharktooth]

Those resize algos are pretty slow. Since quality encodes dont use resizing (anamorphic) i cant see why i sould include such resize algos (assuming you want those resizes in megui).

[Chainmax]

What I meant is to include them in the werewolf comparison you made in post #19. I'd like to see how they'd look compared to the other ones you made.

[Sharktooth]

Ah... btw the werewolf is a picture...

[Chainmax]

I assumed as much, it's just that I'd like to see how EKG and MultiSWAR stack up against the resizers you used in that comparison.

[DewAsmara]

Hi all,
I am not to familiar with those technical explanation, but for me, it shows quite different between DVD 720*576 resolution with upscale at 1280*xxx resolution. But there is something to notice, whether use software or hardware, it's best way to use both of them if have to feed in higher resolution display, example I convert my DVD to DivX HD at 1280*xxx and also set my VGA card at 1440*900 to feed my monitor at 1440*900 resolution. It works quite detail (at least in my eyes) but its only works if I use DVI connection, as its quality down a lot if using analog D Sub. One thing to remember to keep resolution at its original aspect ratio example if 2.35:1 then let it at that way. Changing that only result in terrible pictures. Another is feeding 1280*xxx upscale material from PC to 1280*768 LCD HDTV using D-Sub like my LG only result in the same way with direct connection using DVD player with component cable (almost no difference), as some TV really have good DCDi Faroudja chipset.

Budget DVD upscale player show not much result compare to TV internal upscale engine (use HDMI connection), some even much worst. Buy High End upscale DVD player like Pioneer DV-989AV also not a wise choice if your Plasma/LCD below 40 inch, so think carefully before decide to upscale either hardware or software.

Just fyi, I have been watch upscaling DVD using PS3 in a shop when I am back vacation to my country, it is really shocking, as quality is quite different if running BD from PS3. So don't expect too much with upscale :p

[Single]

I think there is a special case with anamorphic DVD source and the way the video comes to TV.
If you play anamorphic DVD-Video via plain TV analog conection or via any other connection that can transmits only 576 lines per frame to tv that owns higer than 576 horisontal lines resolution then you will have much worse quality despite any smart upscaling or any other enhancers in tv than the case if you use upscaling right after decoding and before transmitting signal to your tv via high resolution connection.

I hope i write this right way in english. :)

[marcobrd]

And here Spatio-temporal interpolation with Motion Vector prectiction and high quality sharpening:

Harry Potter II 720*576*25@1280*720*50
http://multimediacom.free.fr/Video/HPIV_Interpolation.mp4

Interpolation can recreate pixel and can recreate complete frame with MV analyse.

sagittaire (and scharfis_brain), could you please show us some examples of tools and step to get a "Spatio-temporal interpolation with Motion Vector prectiction" resizing? because i think the results you showed in this thread are very impressive. i think this was done with avisinth.. do you use iip?
do you use bobbers like mvbob to get this harry potter example at 50fps? i'm very interested to 576@720 conversion. thank you very much!

[marcobrd]

nothing to say? :)

[scharfis_brain]

i'm very interested to 576@720 conversion.

just
-bob-deinterlace to 50fps using mcbob() (if video is interlaced)
- then resize using gentle limitedsharpen()

every thing else is just not useful

the other 'impressive' (16x120 to 640x480) stuff I presented some time ago was only possible due to the H.E.A.V.Y. aliasing (i.e. flickering) present in the source.
DVD sources generally aren't aliased. In contrary they commonly are severely blurred. So some dehalo-limitedsharpen combo may be appropriate.

[marcobrd]

thank you very much! :thanks:

[DmitryPopov]

Sorry, I've seen this thread too late.

you cant restore details if they were removed.
upsizing a video is not a good idea. what you see on that site is all marketing BS... and the picture examples are indeed fakes.

These words are just an example of not understanding super-resolution. None of the pictures on that site are fakes. Everyone can download the low res files (there are links on the site) and the program, try it and get the same results.

Now I'm proud to say that version 1.5 has been released with much better SR quality.
There is an objective comparion of different upsizing methods, including bicubic, lanczos, lanczos+sharpen and several commercial solutions. Look at PSNR digits and the pictures, get sample files, try it for yourself - everyone's invited to test and see.
http://www.thedeemon.com/articles/video_upsize_methods_comparison.html
My method is definately not Lanczos+sharpen, as it was supposed here, you can clearly see the difference in provided details and sharpness.

Some explanations how super-resolution works are given here:
http://www.thedeemon.com/articles/what_is_super_resolution.html

[Hellworm]

In the special case of video material upscaling with "additional" details can theoretically be done, if the image is in motion. That's because in the case of moving objects there are captured multiple images at different sub-pixel positions of the object. Combining the information of multiple images allows to increase the resolution. There is a program that allows to test this kind of upsizing: ALE (http://auricle.dyndns.org/ALE/)(though not with video material but only with multiple still images)

It really works, but the improvement is not that big. And its slooooow!!

Edit: realized that I answered to the first page and missed the 4 after that.

[eupec]

A simple question: if the source is an *anamorphic* pal dvd 720x576 (video aspect ratio ~1.81 excluding black borders) and I "resize" via avisynth filtering to an upper resolution, lets' say 856x472 it's a mistake because I add data (by interpolation or other way) which wasn't existent before resizing OR it makes sense?

Thank you :)
Eupec

[*.mp4 guy]

Unless the screen you are displaying it on is 856x472, it was a mistake.

[eupec]

Unless the screen you are displaying it on is 856x472, it was a mistake.

Thank you for reply:)
Suppose I've got a screen at this resolution; it's because I feed the screen with a video at its own resolution, so the screen don't have to scale internally?

Eupec

[*.mp4 guy]

yes, essentially; all video will be rescaled to the resolution of your screen on playback, in some cases (mostly lcd televisions) this is done by the screen, however it is usually done by the video card. Both ATI and nVidia use a simple bicubic kernel, so its pretty easy to get much better looking video by, for example, using ffdshow to rescale the video to your screens resolution with high tap lanczos, on playback. Obviously if your source is a blury mess or severly frequency limited already it doesn't matter what type of resizing you use.

[patrick_]

Another question: Can you improve image quality doing only horizontal upscaling?

The source in question is a DVD with anime content. It's pal so has a resolucion of 720x576 (anamorphic). We're compressing this video using x264 into a MP4.
Will results improve when upsizing to 1024x576? Or is it better to maintain the original resolution and to set PAR within the MP4.

In case upsizing is better, what algoritm should be used?

All programs support PAR? MPC does.

Thanks in advance.

[Dark Shikari]

Another question: Can you improve image quality doing only horizontal upscaling?

The source in question is a DVD with anime content. It's pal so has a resolucion of 720x576 (anamorphic). We're compressing this video using x264 into a MP4.
Will results improve when upsizing to 1024x576? Or is it better to maintain the original resolution and to set PAR within the MP4.

In case upsizing is better, what algoritm should be used?

All programs support PAR? MPC does.

Thanks in advance.
Since upscaling it will require more bitrate, I would think the extra bitrate given to the original file, without upscaling, would give more of a quality improvement than the very slight improvement given by the Lanczos upscaling.

[*.mp4 guy]

Since upscaling it will require more bitrate, I would think the extra bitrate given to the original file, without upscaling, would give more of a quality improvement than the very slight improvement given by the Lanczos upscaling.

Thats only true, If all you use is lanczos, or if you aren't resizing to your screens exact resolution, for example:

bicubic-neutral (http://img221.imageshack.us/img221/1492/bicubicwv7.png)
fancy-slow (http://img218.imageshack.us/img218/8858/slowof5.png)

Obviously personal preference, and specific usage conditions play a big role in determining if slower-then-realtime interpolation is a good idea, but it can be much better looking then on-playback interpolation.

[patrick_]

Thanks for responding.

The video will not be reproduced on a 1024x768 screen. So whatever method will be used, it must be resized when reproducing.
So will pre-coding resizing results in a better result? It means that the video will be resized twice. Without pre-coding resizing it will only be resized once.

[*.mp4 guy]

Well, You still haven't iven enough information to answer that question adequately. The answer depends upon, 1- how good the resizer you are going to use is (if your just using lanczos or spline, your better of not ressizing) 2- how much compression you need to achieve (if you are, for instance trying to fit a full length movie on a cd, upsizing is always a bad idea).

[patrick_]

Interesting point. Fitting a full length movie on a cd isn't a problem with superb quality. You can easilly convert Frequency to a 500MB H.264 with hardly any loss in quality. <those 500MB is from memory, maybe it's 550 don't remember for sure
Of course it depends a lot on the movie, you can't do that with most movies.

I just finished converting Akira and it's only 877MB. It's the best version I know of on internet. I must admit that I did take out almost all noise (without losing details, and restored sharpness) - with noise size would be 1280MB. Resultion is 720x400. Source is progresive PAL (but for some reason the DVD said it's interlaced)

I only use x264 with maximum processing settings. (with StaxRip)

My question isn't related directly to the conversions I do. It's just about a simple discusion with another fansub. They upsize to 1024x576 to maintain the 576 lines, but I don't know what filters they use (unlike I do, they don't share their scripts). They use 233MB with 5.1 sound for about 23 minutes.

In my conversions I do never try to fit a video on a cd or a season on a DVD. I use a constant quality (55% - StaxRip).
That's also the second point of our discusion. They (like everyone over here) convert all episodes to the same amount of MBs. It's obvious that episodes with less movement don't need the same bitrate, so they're just wasting space. BTW I think most English-language fansubs do the same thing.

Getting back to your response, can you give me the name of some real good resizers? (including comercial ones). Also I'm interested in a (comercial) deinterlacer with a quality similar to the Faroudja of my proyector.

[Trahald]

I've been doing dvd->hd recently (480p->1080p) i have just been doing 4tap lanczos. the result is better than the on the fly resizer, but obviously not as good as true hd.

@*.mp4 guy
What resizing routine did you use for your fancy-slow png. what was the source size?

[*.mp4 guy]

@ patrick
Try mcbob for a deinterlacer, its free, but its the best one I know of, better then Faroudja imo, but very slow. You might also want to brush up on the forum rules.

@Trahald
The source was the Animatix dvd (NTSC), so 704*360 ( I think I downsized to 704*304, then upsized from there for the slow resizer, to save time; I don't know becuase I processed a short clip form the animatrix a while ago, and the screenshot is from that) the bicubic upscale was run on the full 704*360px (I just did the bicubic upscale today, to give a reference). The resizer I used is an unrealeased version of lanczosmtplus, an avisynth script based high quality upsizer I wrote.

[patrick_]

Thanks I'll have a look at mcbob (right now I was using Yadif).
But.... can't seem to find a download link. I've found some v0.3 avs, but that's all. Nothing else. I guess there must be some oficial download. Can you please give me a link?

[Didée]

Yes indeed, a dingy Avisynth script, that's all.


"Official" Link to MCBob v0.3c: >clickme (http://home.arcor.de/dhanselmann/_stuff/MCBob_v03c.avs)<

"Official" Thread for MCBob: >clickme (http://forum.doom9.org/showthread.php?t=124676)<

Inofficial "how-MCBob-was-born" thread: >clickme (http://forum.doom9.org/showthread.php?t=117025)<


In case that smells like rummage - feel free to remit plenty of $$$ to me. The results will appear much more worthy after that. :D ;)

[patrick_]

Thank you for the links.

BTW ever thought of creating a zip with all filters requiered? (just like mvbob). Would be much easier than spending a lot of time searching for all required plug-ins. (or at least some comments in the avs with directions where to download every plugin).

As you're the writer, I have another question, Can I mirror the download? (actually I don't want to mirror it, but have it within a zip as explained before).

[patrick_]

I noticed that mvbob dobles the framerate. Does mcbob do the same thing? If so, what's the best way to reduce the framerate without losing all the effors done by the deinterlacer.

Thought about ChangeFPS, but I'm not sure about how much quality will be lost.