[MPlayer-users] Comparison of different software scaler types

[Tuukka Toivonen]

On Wed, 15 Oct 2003, Matthias Wieser wrote:

>> There is no correct (much less no unique correct) way to scale images
>> down.

>I think, there is one.
>A good downscaler should try to
> -preserve as much as possible details
>- not introduce artifacts or noise (as -sws 8 does)
>After scaling up, the difference to the original imige should be minimal.

Since you don't specify upscaler, you're actually talking about
compression, not scaling. "make image smaller so that when restored
to the original size there is as little differences as possible"
is nothing else except compression.

The point is that the image should look as much to the original
when scaled down as possible. But since human preferences and
display devices vary, there isn't a single best algorithm or
even a meter which could be used to judge "goodness" (except
human viewers... but even then it's display device dependent).

Many people think about signal processing and would say that an
ideal lowpass filter with infinite extend would be best, but it's
not true--human eye is nonlinear so linear signal processing is
not subjectively that best alternative. There isn't really one.

>> If you want a better test, you should use MPlayer to scale the images
>> *UP* with different filters, and then scale them back down and compare
>> to the original.

And what method should be used for scaling them back?
Btw, I'd be surprised if a filter used for scaling up would produce
equally good results for scaling down. And the scaling ratio matters much,
nearest neighbour scaler isn't that bad when scaling exactly 2x larger.
1.5x and it's much worse.

>But do you know, why -sws 8 is so broken? Sinc should produce very good
>quality (see

I don't know about colors but at least the ringing was pretty bad
when I tried it.