Noise filter before or after resize IF up-sizing ? What IF down-sizing?

[PiXuS]

I found these posts:

http://forum.doom9.org/showthread.php?s=&threadid=51862
http://forum.doom9.org/showthread.php?s=&threadid=36344
http://forum.doom9.org/showthread.php?s=&threadid=64253

But in none of them I found specific mention of whether (not wether!) putting the noise filter (whatever noise filter) before the resize filter was still appropriate (quality wise) when UP-sizing the video resolution.

e.g.


SuperDupperNoiseFilter()
SharpResize(10240,7680)


or


SharpResize(10240,7680)
SuperDupperNoiseFilter()


Anyone?

Thanks!

[kingmob]

I'd always put the filter before any resize. For downsizing, you'll get far better results and for upsizing it could only hurt you to do it afterwards.

[Manao]

It depends...

After downsizing, the denoising is faster. At the same settings, it is usually stronger too, so you have to take that into account.

After upsizing, it's slower, and the denoising is weaker ( but allows a more subtile denoising ).

Before upsizing, it slightly avoids upszing noise, but in order to get a fully noise reduction, you often have to blur too much the picture, which means you'll need a sharpening filter after resizing.

For upsizing, I would do temporal denoising -> upsizing -> spatial denoising -> sharpening. But I scarcely upsize my videos, and some may disagree with me.

[PiXuS]

Originally posted by Manao
It depends...

After downsizing, the denoising is faster. At the same settings, it is usually stronger too, so you have to take that into account.

After upsizing, it's slower, and the denoising is weaker ( but allows a more subtile denoising ).

Before upsizing, it slightly avoids upszing noise, but in order to get a fully noise reduction, you often have to blur too much the picture, which means you'll need a sharpening filter after resizing.

For upsizing, I would do temporal denoising -> upsizing -> spatial denoising -> sharpening. But I scarcely upsize my videos, and some may disagree with me.

Thanks Manao. In your filter chain "temporal denoising -> upsizing -> spatial denoising -> sharpening" would you use Unfilter(5,5) (or some other positive values) to sharpen? If you omitted the spatial denoiser, would you instead use Unfilter(-5,-5) to unsharpen the picture (given the upsizing was done using Lanczos)?

Again, thanks for your prompt answer.

[IanB]

@PiXuS,

Generally from information theory, when upsizing there is no more real information available. So filtering after the resize shouldn't be any better, just slower.(Manao) After upsizing, it's slower, and the denoising is weaker (but allows a more subtile denoising).That said however, most filter don't use highest precision arithmetic and do deal with 8 bit data on input and output, this leads to small inaccuracies in the result. Sometimes, not often, these are visible. By applying such a filter after upsizing some such small imprecisions may be mitigated (However by increasing the precision of the original filter algorithm apropriatly this point becomes mote).

In general filter before upsizing, but there will be examples where upsizing first can ease the mathematical precision requirement of a filter and yield better results (I don't expect there to be many such cases).

IanB

[PiXuS]

Originally posted by IanB
@PiXuS,

Generally from information theory, when upsizing there is no more real information available. So filtering after the resize shouldn't be any better, just slower.That said however, most filter don't use highest precision arithmetic and do deal with 8 bit data on input and output, this leads to small inaccuracies in the result. Sometimes, not often, these are visible. By applying such a filter after upsizing some such small imprecisions may be mitigated (However by increasing the precision of the original filter algorithm apropriatly this point becomes mote).

In general filter before upsizing, but there will be examples where upsizing first can ease the mathematical precision requirement of a filter and yield better results (I don't expect there to be many such cases).

IanB

Thanks. Very informative.

[Didée]

IanB:

Generally you're right, not much to say. However there are a few instances where "supersampled" filtering really helps. For example in sources with high sharpness, on scenes with very slow movements there is often some kind of shimmering (pixel aliasing due to the high sharpness) that can't be catched by temporal smoothing with reasonable thresholds. In such cases, simulating subpel precision by supersampling can give better results.

Example: chapter 27 of LOTR-TTT SEE. After ca. 15 seconds into that chapter, there's a slow panning on a landscape, with noticeable shimmering in the center of the frame. Try to catch that shimmering with whatever filter applied in original resolution: it's hopeless.

[Teegedeck]

Sheesh, why don't people with similar dedication work in the industry? It is damn ridiculous that you should have to filter out odd effects yourself. I don't have time for such stuff. :o