Influence of Scaling Algorithms on EE

by Bjoern Roy , September 10th 2003

Different scaling algorithms induce different amounts or ringing.

If a screenshot of a DVD is scaled with one of those algorithms, another layer of ringing is induced 'on top' of the EE/ringing that is encoded on the DVD. Is it then still valid to even use 'scaled' screenshots to in reviews to analyze EE on DVD?

The scaling algorithms in external scalers or HTPCs also induce such another layer of ringing. Since DVDs are mostly plagued by source encoded ringing/EE anyway, wouldn't it be best choose the scaling algorithm with the least amount of ringing?

Both valid concerns that i am trying to tackle in this editorial. The question why to scale at all is a different topic that i won't touch today....

 

I will try to show that this 'additional' ringing layer, that is introduced through scaling, is basically irrelevant under real world situations, and thus any concerns that DVD reviewers use these algorithms, is only 'theoretically' valid, and that it is as such more of a moot point really.

I prepared a comparison picture that demonstrates the effect that different scaling algorithms have on various amounts of source inherent ringing on DVD transfers. For the comparison, i upscaled frames from 6 different DVDs from NTSC SD resolution (720x480) to HD (1920x1080), mimicing a high-end HT application (e.g. 9" CRT with scaler), or a DVD review that uses upscaled screenshots to show deficiencies (like i do here on my site). I then used equally sized 200x225 pixel big crops from these upscaled frames to focus on 'critical' areas, where the level of ringing is clearly visible.

Here is the picture. Its probably best to open it in a different window:


[Click image to enlarge]

 


From left to right, the amount of DVD inherent ringing increases considerably:

1.) AVIA
No ringing at all. These are black lines on a gray background.

2.) Fast and the Furious
Very high level of detail, yet basically no ringing. Great.

3.) X-Men
A tad more ringing. Still very good.

4.) Unbreakable
Quite thick and strong halos, equally in both dimensions. Bad.

5.) Star Wars: The Phantom Manace
A real classic. Ugh!

6.) Die Hard 3, first edition
Sensational.


I used 4 different scaling algorithms to upscale these 6 samples. They too differ considerably in regard to the amount of ringing they induce, only the first one doesn't induce any ringing at all. Why would someone choose any other scaling algorithms than the first one then, you might ask?

Well, its not that easy. Going into more detail on scaling is on my to-do list, but thats a broad and complex topic. But it should be suffice to say that the scaling algorithms are always a tradeoff between 4 characteristics: ringing, response/detail, postaliasing and computational demand. Bilinear, which is the filter that inherently doen't induce any ringing of its own, is also very easy to compute, but isn't optimal postaliasing wise and its transfered response is the most soft/blurry of the commonly used scaling algorithms.

So for HT scaling applications, bilinear filtering is a rather bad choice. For DVD reviews, it would do, but as is my very intend to demonstrate in this post, the ringing induced through the other filters does't really affect the screenshots sufficiently to invalidate their usefulness.

From top to bottom, the amount of scaling induced ringing increases:

[Note: To inspect these differences in ringing, its advised to focus on the leftmost AVIA pattern, since its the only one without any source inherent ringing. Lateron, i will discuss the effect this has on the actual movie samples!]

1.) Photoshop Bilinear Filter
The topmost row is upscaled with Photoshop's bilinear filter. As already mentioned Bilinear scaling doesn't induce any ringing of its own, as can be seen on the AVIA crop on the left. Its incapable of transfering the frequency response properly, though, resulting in a blurry picture (e.g. Van Diesel's face).

2.) Bell Filter
This is the filter that i use for my reviews. It does induce the very slightest amount of ringing (again, see AVIA pattern), so little, that most probably wouldn't notice. Response is a bit better with this filter.

3.) Photoshop Bicubic Filter
This is probably the most common upscaling algorithm. Graphic cards like Radeon/Geforce are using this since years, some external scalers as well. Resampling in Photoshop and other tools is very commonly done with it. And its probably also used by most of the other DVD review sites that use scaled screenshots like i do. The reason for its popularity is the good 'balance' between the 4 characteristics that i mentioned above.

If you inspect the AVIA pattern, you can see that it has indeed quite a bit more ringing than the earlier 2 filters, but its still rather simple to compute (at least with reasonable support/neighborhood) and has very good response.

4.) Lanczos-windowed Sinc Filter
The family of Sinc filters are 'ideal' in the sense that they theoretically transfer the response flat. In its pure form, a sinc filter would have infinite support, thus being uncomputable. Even with restricted support, and using a Lanczos-window is one such restriction, they are very compute intense.

The price you pay for the 'perfect' response is considerable ringing.


Now, if you only look at the AVIA pattern, you could think that the ringing that is induced through the different filters is rather severe, especially in case of the Lanczos filter. Yet, if you look at the second sample (Vin Diesel), you'll see that there is not that much if any actual difference in ringing.

The tree in the top left for example is a prime 'potential' candidate for ringing. Heck, on 90% of all DVDs, this would be a mess no matter how you slice it! Yet, even the Lanczos filter doesn't manage to turn this into anything but smooth, ringing-free bliss. [Hey, i am not talking about the movie here :-) ]

Why is it that the AVIA pattern shows the difference in ringing so blunt, yet that tree doesn't? The reason is, that the the resampling filters mostly produce ringing from frequencies close to the passband limit (e.g. 6.75Mhz in case of DVD). The pixel perfect lines in the AVIA pattern HAVE response flat to that limit. Thus, they induce heavy ringing.

All 'real' DVDs are heavily filtered, though. In the horizontal direction, the Superbit titles are the closest to being unfiltered, although at the cost of aliasing. Vertically, the Superbits are just as much filtered as any other title to minimize interlace flicker. If specially tailored for progressive display, the Superbit concept could also be used vertically (opening up the filter), yielding yet another gain in picture detail. 'Superbit squared'

Anyway, because DVDs are so heavily filtered, they don't contain much, if any, response close to the passband limit and are thus less likely to cause much scaling related ringing. So basically the primary tradeoff of scaling algorithms: response/detail vs ringing is slightly tilted with 'real world DVDs'.. The negative aspect ringing, is not as much of a problem. The positive aspect, higher response, preveils.

So what you basically see when you inspect all the 5 movie samples, is that the different filters don't really 'add' any ringing that isn't there. What it does, is make the ringing that is there, 'stand out' more, since the transfered frequency response is higher towards higher freuencies and 'ringing' is mostly higher frequencies.

No matter which row of samples you watch, its always apparent that Furious has none, Unbreakable has considerable, and the last 2 have excessive ringing. The 'relationship' among these DVDs in regard to ringing is always the same.

And thats the important thing here. ALL upscaled screenshot samples i have ever seen posted, always really really actually demonstrated the deficiencies that are present ON THE DISC. No matter what upsampling filter used. And since bicubic isn't even the filter that introduces the most ringing of its own, yet its the most used one, makes this a complete non-issue.

Again, the reason you see a bit more ringing for example in the Lanczos Unbreakable sample is NOT because it 'adds' that ringing. Its because the transfered frequency response is much better with that filter, thus those higher frequencies of which the ringing consists, are less subdued, thus more visible. So the bilinear filter 'blurrs away' some of that ringing. The problem is, that 'real detail' in that freqency range are blurred away as well.

I laid out my case, honorable jury. I vote 'not guilty' :-)

Best regards
Bjoern Roy

 

 

Back to Editorial Main Page

WEBCounter by GOWEB WEBCounter by GOWEB