Reducing edge noise in animation SVCD

[kinneera]

Anyone have some good tricks for reducing the amount of edge noise that is so visible when encoding animation at SVCD bitrates? The source is nothing great (cable), but it's still pretty bad to look at on a monitor, even after noise reduction.

[Shabubu]

Try adding temporal of spatial smoothers b4 encoding. And remember to set your Quantize Matrix in TMPGEnc to animation. Lukes video guide (can't remember the address) deals pretty much exclusively with animation encoding and capping. Theres a link in these forums somewhere.

[KoalaBear]

Kinneera, I'm only offering you this suggestion because I know that you know what you're doing and you won't hound me for support afterwards (grin).

This is one of the few instances where changing the GOP (and possibly the quantization matrix) will help you. Dump as many Bs as possible for Ps, and shorten the GOP if necessary to accomodate the increased bitrate.

Animation is extraordinarily difficult for MPEG, because it's exactly the opposite of what it was intended to encode. Even DVDs look like crap when played back on a PC; you can actually see the picture "pulsing" twice a second at the appearance of every I-frame. The errors accumulate that much, that quickly.

That's why fewer B-frames are a good idea; P-frames don't propagate as many errors, and as long as you can control the overall bitrate you'll be ahead of the game.

But remember that even the finest animation, even on DVD, looks like crap on anything other than a TV. It's the absolute worse case for MPEG. The best you can hope for is lack of visibility versus lack of persistence.

[kinneera]

Thanks for the suggestions. I've read some about the techniques MPEG employs for compression, and I was just wondering why animation is the absolute worst case? It seems like the large blocks of constant color would would make it easier for it to do I frame compression and find motion vectors. Just curious.

[KoalaBear]

Mostly because of DCT. Discrete Cosine Transform was chosen over Fast Fourier Transform because it approximates lines not exactly, but well enough. See here for details.