[资料]Convolution3d[英]

dgwxx

File : Convolution3d.txt

Copyright © 2002 S閎astien LUCAS. All rights reserved.
babas.lucas@laposte.net

This file is subject to the terms of the GNU General Public License as
published by the Free Software Foundation. A copy of this license is
included with this software distribution in the file COPYING. If you
do not have a copy, you may obtain a copy by writing to the Free
Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.

This software is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details


*****************************************************************************************

Convolution3D is an avisynth filter that will apply a 3D convolution to all pixel of a frame.
for all future release see at http://hellninjacommando.com/con3d/

1 - How to use it :
Convolution3d (matrix=0, ythresh=3, cthresh=4, t_ythresh=3, t_cthresh=4, influence=3, debug=0)

Matrix choice :
0 : original matrix :
1 2 1 2 4 2 1 2 1
2 4 1 4 8 4 2 4 1
1 2 1 2 4 2 1 2 1
This matrix is useful for normal movie (not anime) because it keep more details
1 : bb idea of full 1 matrix (great idea)
1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1
This matrix is much usefull with animes or bad quality sources because it blur a little more (so removing more noise)

Temporal influence :
It's used especially to speed up a little this filter and to avoid using temporal informations when not needed
(scene change, fade, ...)
I first build a limit = Temporal Luma Threshold * Temporal influence
For each 2 pixel computed (due to MMX, 2 pixel at the same time), I first check this :
if
(Abs (Y0 - Y0[Previous frame]) +
Abs (Y0 - Y0[Next frame]) +
Abs (Y1 - Y1[Previous frame]) +
Abs (Y1 - Y1[Next frame])) > limit
then
do Spacial work (only 3*3 matrix)
Else
do Spacial and Temporal work (3*3*3 matrix)
The lower it is -> the faster will be the filter but compressibility should be lower
The higher it is -> the slower will be the filter but compressibility should be higher
if temporal influence is set to -1 then only spatial work is done (high speed).
This parameter is a float.

2 - Parameters sample

I build the following presets to make things easier :
Convolution3d (preset="movieHQ") // Movie Hi Quality (good DVD source)
is an alias for Convolution3D (0, 3, 4, 3, 4, 2.8, 0)
Convolution3d (preset="movieLQ") // Movie Low Quality (noisy DVD source)
is an alias for Convolution3D (0, 6, 10, 6, 8, 2.8, 0)
Convolution3d (preset="animeHQ") // Anime Hi Quality (good DVD source)
is an alias for Convolution3D (0, 6, 12, 6, 8, 2.8, 0)
Convolution3d (preset="animeLQ") // Anime Low Quality (noisy DVD source)
is an alias for Convolution3D (1, 8, 16, 8, 8, 2.8, 0)
Convolution3d (preset="animeBQ") // Anime Bad Quality (???)
is an alias for Convolution3D (1, 12, 22, 8, 8, 2.8, 0)
Convolution3d (preset="vhsBQ") // VHS capture Bad Quality (???)
is an alias for Convolution3D (0, 32, 128, 16, 64, 10, 0)



I had to test Convolution3d with bad quality TV capture and in this case
you'll have to higher especially the chroma tresholds (causing some ghosting but the overall quality seems to be better).
I personnaly use these parameters :
Convolution3D (0, 32, 128, 32, 128, 10, 0)


The thresholds of Convolution3d are only here to take care of edges and scene change. You can increase the spatial one (especially the chroma threshold) but stop as soon as you see some blurring around the edges (if you want quality). With the settings proposed you shouldn't have this problem.
The Temporal one should be left below 10 to avoid ghosting.
You should especially take care of the threshold with matrix 1, because with this matrix the current frame has less weight so it's easier to have ghosting.

You can find some informations about how it works in : http://forum.doom9.org/showthread.php?s=&threadid=29829
You can also find a french tutorial in : http://forum.media-video.com/viewtopic.php?t=7682

3 - Current limitations or known problems : .
- Work only with YUV2, CHECKED.
- require a Integer SSE capable CPU (no PII and K6-II), CHECKED.

WARNING : it's slow, I know it and I try to make it faster so don't rush me about it, thanks in advance.

4 - Credits

Thanks to
bb for the original idea and a lot of tests
iago, Koepi and TheReal for real full length movie (or capture) tests
Tom Barry, Dividee and Sh0dan for their usefull technicals informations or ideas
Ctrl-Alt-Suppr for a french tutorial
Defiler for hosting Convolution3d
all Convolution3D users


Vlad59 (babas.lucas@laposte.net)