Smooth anime for compressibility

[bb]

I've just had another idea for a filter, maybe someone kicks in...

The sharp edges eat up the most bits in DCT based codecs like DivX, XviD, MPEG2, etc. And here's my idea:

Create a filter which detects the sharp edges and smooth these areas only.

What do you think?


bb

[Acaila]

But the sharp egdes are exactly what determines the detail. Blur the edges and you'll lose detail real fast (or at least the illusion of detail). That is why all filters try NOT to blur sharp edges.
It's the even colored planes that become ugly when artifacts rise up, and smoothing those already brings a nice compression boost (and even enhance quality sometimes) with existing filters. If you start blurring the edges you'll see that you have to use very very mild settings before you blur too much, in which case the compression boost will be quite small.

[sh0dan]

I'll also have to support Acaila here. I'd rather have a few more artifacts than risk loosing details. Also ffdshow has great postprocessing that gets rid of most of these problematic areas. So for once I'm more to post-proccessing than pre-processing! ;)

[bb]

Blurring the sharp edges would be useful for the compression IMO. If you prefer post-processing, why not use a sharpening filter for post-processing? I think it's worth a try...

bb

[DJ Bobo]

@ bb
You idea is nice, but I think it's already done: BilinearResize ;)
In fact, BilinearResize blurs the edges a bit, to an acceptable level, which improves compressibility pretty well without sacrifice of details.
I use BilinearResize a lot on Anime and I'm very satisfied with it.

[bb]

Yes, DJ Bobo, but it blurs the rest as well. Is there nobody who supports my idea?

bb

[dividee]

You can achieve that effect using layer:
use Sobel edges detection, blur the edges, apply it as a mask to a blurred copy of the vid and use layer add operation to mix with an original copy of the vid.
Look in the "averaging two captures for noise reduction" thread. On page 2 poptones posted a script that does almost that, but ony blurs stationnary edges. You could start with that one and modify it. It's quite a challenge to understand it at first, but it will allow you to understand the power behind layer and convolution.