DivX 6.6 extreme noise reduction settings?

[evilclive]

I do most of my encoding in DivX 6.6, at CQ=4.

Am I right in thinking that the noise reduction feature in DivX 6.6 is a smoothing pre-processor, and therefore that setting the noise reduction to extreme, solely to reduce the size of the file, will result in a lower picture quality?

How bad does the picture have to be, to benefit from extreme noise reduction?

Assuming it's a reasonably good analogue terrestrial recording, with the usual amount of low-level screen noise, is it worth me attempting encodes at CQ=4.5, light noise reduction, and CQ=4.0, extreme noise reduction, and comparing the results? (The recording is 90 minutes long, and at CQ=4, doesn't fit on a CDR unless I use extreme noise reduction.)

How reliable is the auto-detect noise reduction setting at optimising the results?

[olnima]

...setting the noise reduction to extreme, solely to reduce the size of the file, will result in a lower picture quality?...

...depends on YOUR definition of quality


...How bad does the picture have to be, to benefit from extreme noise reduction?...

a bit bad or maybe a bit more bad but not too bad, or, to say it with numbers: 3-4, 5 is too bad...
Give me any idea how to answer this question seriously...

...Assuming it's a reasonably good analogue terrestrial recording, with the usual amount of low-level screen noise, is it worth me attempting encodes at CQ=4.5, light noise reduction, and CQ=4.0, extreme noise reduction, and comparing the results? (The recording is 90 minutes long, and at CQ=4, doesn't fit on a CDR unless I use extreme noise reduction.)

How reliable is the auto-detect noise reduction setting at optimising the results?

Assuming You have an internet-connection and therefore the possibility to download divx, how reliable is it testing all these things for yourself?

I use the noise-reduction at its max. level but I really can't tell if You would like MY results or not... :)


Olnima


//edit: wow, I just saw that You DO have divx 6.6 already... double wow!

[evilclive]

...depends on YOUR definition of quality

I wasn't really asking the usual unanswerable, subjective question - I was actually after some documentation.

After a bit of searching, I found a reference guide at http://download.divx.com/divx/DivXUserGuide521-en.exe

Source pre-processing applies spatial and temporal filtering prior to encoding in order to reduce noise in the source video.

Four presets are defined allowing control of the filter intensity. These are light, normal, strong and extreme.

Both light and normal presets are very carefully balanced and neither should cause visible degradation to the video.

Sources with heavy noise present may require pre-processing using the strong or extreme presets. The presets may introduce a visible smoothing effect in the video, therefore when using these two presets it may be advisable to process short sample clips under each to find the best balance between noise reduction and smoothing.

There you have it. Applying too much noise reduction will reduce the quality of your image, irrespective of the Quality setting of your transcode. This applies especially if, like me, you use CQ, rather than worrying about fitting the file nicely on a CDR.

I use the noise-reduction at its max. level but I really can't tell if You would like MY results or not... :)


It looks like you've fallen into the same trap that I did.

Incidentally, the Automatic setting seems to adjust the noise reduction dynamically, according to need. I had a noisy HuffyUV file that I converted to MPEG-1 (with noise reduction) years ago, because I couldn't afford the storage requirements of the original. Now the noise is minimal, but the MPEG-1 artifacts are somewhat irritating, to say the least.

This was a custom MPEG-1 setting, using CQ=85 VBR with an effectively unlimited maximum bit rate, and so it won't play on a standard DVD player. To make it viewable on a TV, I'm converting it again, to DivX 6.6.

At CQ=3.0, the file sizes of just the video elemental stream are:

No noise reduction: 894 MB
Light noise reduction: 774 MB
Automatic noise reduction: 891 MB.

Even without attempting a subjective evaluation of the picture quality, I can quantify the amount of noise reduction by the reduction in file size. Presumably, almost all the noise has already been removed from the source MPEG-1 file, and a second smoothing operation would be more deleterious than usual to the picture quality.

The automatic noise reduction seems, on this evidence, to select the minimum amount of noise reduction consistent with improving the picture. Whether it's reliable enough to be trusted blindly is a matter for anecdotal evidence.

[olnima]

...There you have it. Applying too much noise reduction will reduce the quality of your image, irrespective of the Quality setting of your transcode.
Every filter reduces the quality objectivly but the only thing that interests me is, if I like the results or not...

...It looks like you've fallen into the same trap that I did...

...no, not a trap. For me this is the setting from which I get MY best results depending on MY eyes, MY brain, quality and filesize. I always use CQ (capturing directly into divx) and I found MY best settings not with asking in this forum but with three other things: trying, trying, trying...
The only thing I want to say is that nobody can watch with your eyes/brain so nobody knows what You like or not. Is a sharp and blocky picture better than a smooth one?
Answer: It depends...

And if You have already divx why don't You try and find YOUR best settings. The only thing I can say is that the automatic mode in noise reduction needs more CPU Power than the other ones and that it doesn't make me happy. But maybe You get another result. Try and error is the best and at least the only thing giving You YOUR answer. ;)

Good luck,
Olnima

P.S.:

...therefore...it may be advisable to process short sample clips under each to find the best balance between noise reduction and smoothing...
That's what I meant ;)

[evilclive]

I've just done some sample encodes.

My source file is a 90 minute MPEG-1 at 352x288, 2400kbps CBR that was recorded from an analogue terrestrial TV signal with an original All in Wonder Radeon some years ago. The amount of broadcast noise in the unfiltered MPEG-1 source file is such that I would instinctively want to use at least some noise reduction.

Zero noise reduction - -- (CQ=3) 1164MB -- (CQ=4) 933MB
Auto-detect noise red. -- (CQ=3) 1144MB -- (CQ=4) 915MB
Light noise reduction - -- (CQ=3) 964MB --- (CQ=4) 761MB
Normal noise reduction -- (CQ=3) 892MB --- (CQ=4) 704MB

These files contain the VBR video stream only; if I wasn't using the fast preset, the files would be a little smaller.

Whatever the auto-detect noise reduction feature does, it doesn't seem to be choosing an optimum noise reduction level for CQ transcoding. (This is hardly surprising, as a one-pass transcode cannot feed its output back into the pre-processor input.)

My impression of the output is that the blocky, lossy nature of the MPEG-1 source is noticeably magnified by the transcode, so I prefer CQ=3 for this purpose. For a transcode of a Freeview recording (4Mbps MPEG-2, free from white noise), I find CQ=4 perfectly adequate.

I still need to play with the NR settings, because I still don't see what (other than 25% of the file size) I'm losing by turning NR on.


Edit (midnight): I think the auto-detect NR feature is more for use with limited-bit-rate (CBR, multipass ABR) settings. It also seems that the noise filter is throwing away very little other than the source file's MPEG artifacts; and a low-pass filter is just about as crucial for DivX as it is for MP3 audio encoding.