[XviD-devel] image preprocessing
Marc FD
xvid-devel@xvid.org
Thu, 8 Aug 2002 18:07:13 +0200
This is a multi-part message in MIME format.
------=_NextPart_000_0025_01C23F06.6C002E20
Content-Type: text/plain;
charset="iso-8859-1"
Content-Transfer-Encoding: quoted-printable
I experimented with some preprocessing filters again yesterday. I used =
a 3x3 averaging filter kernel as well as a symmetric 7-tap low-pass one. =
Results are similar with both filters: The PSNR values between the =
unfiltered source and the filtered reconstructed images are much lower =
than without filtering, but that's of course no surprise: One can't =
expect that such prefiltering brings you closer to the original =
(PSNR-wise).
However the PSNR values between filtered source and filtered =
reconstructed images are (depending on the strength of the filters) =
significantly better than their unfiltered equivalents. This proves that =
such prefilters help to avoid coding artifacts, so as long as the =
filtered image remains pleasing to the human eye (and I'd say it does), =
even such a simple prefiltering step can be an improvement.
sounds very interessing.
maybe there good improvements for XviD here (ie on anime ?)
I would really want to help this project :)
btw: does someone know what sort of filter divx5 uses? I did a divx5 =
encode with preprocessing yesterday and the result looked somewhat =
strange...
i can just say what i experienced as an old DivX 5.x user (converted =
to XviD :) )
i think DivX doesn't use "real" preprocessing. or the reconstruction =
filters are very bad,
because it just gives the same results than some good cleaners with =
strong settings.
it seems to be only some washing out option (just destroy extremes dct =
coefs ??)
PS : what's a "symmetric 7-tap low-pass one" ??
MarcFD marc.fd@libertysurf.fr
------=_NextPart_000_0025_01C23F06.6C002E20
Content-Type: text/html;
charset="iso-8859-1"
Content-Transfer-Encoding: quoted-printable
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN">
<HTML><HEAD>
<META http-equiv=3DContent-Type content=3D"text/html; =
charset=3Diso-8859-1">
<META content=3D"MSHTML 6.00.2600.0" name=3DGENERATOR>
<STYLE></STYLE>
</HEAD>
<BODY bgColor=3D#ffffff>
<BLOCKQUOTE dir=3Dltr=20
style=3D"PADDING-RIGHT: 0px; PADDING-LEFT: 5px; MARGIN-LEFT: 5px; =
BORDER-LEFT: #000000 2px solid; MARGIN-RIGHT: 0px">
<DIV><FONT face=3DArial size=3D2>I experimented with some =
preprocessing filters=20
again yesterday. I used a 3x3 averaging filter kernel as well =
as a=20
symmetric 7-tap low-pass one. Results are similar with both =
filters: The=20
PSNR values between the unfiltered source and the filtered =
reconstructed=20
images are much lower than without filtering, but that's of =
course no=20
surprise: One can't expect that such prefiltering brings you closer to =
the=20
original (PSNR-wise).</FONT></DIV>
<DIV><FONT face=3DArial size=3D2></FONT> </DIV>
<DIV><FONT face=3DArial size=3D2>However the PSNR values between =
filtered source=20
and filtered reconstructed images are (depending on the strength of =
the=20
filters) significantly better than their unfiltered equivalents. This =
proves=20
that such prefilters help to avoid coding artifacts, so as long as the =
filtered image remains pleasing to the human eye (and I'd say it =
does), even=20
such a simple prefiltering step can be an improvement.</FONT></DIV>
<DIV><FONT size=3D2></FONT> </DIV>
<DIV>
<DIV><FONT size=3D2><STRONG>sounds very =
interessing.</STRONG></FONT></DIV>
<DIV><STRONG><FONT size=3D2>maybe there good improvements for XviD =
here=20
</FONT><FONT size=3D2>(ie on anime ?)</FONT></STRONG></DIV>
<DIV><FONT size=3D2><STRONG>I would really want to help this project=20
:)</STRONG></FONT></DIV></DIV>
<DIV><FONT face=3DArial size=3D2></FONT> </DIV>
<DIV><FONT face=3DArial size=3D2>btw: does someone know what sort of =
filter divx5=20
uses? I did a divx5 encode with preprocessing yesterday and the result =
looked=20
somewhat strange...</FONT></DIV>
<DIV><FONT face=3DArial size=3D2></FONT> </DIV>
<DIV><FONT size=3D2><STRONG>i can just say what i experienced as an =
old DivX=20
5.x user (converted to XviD :) )</STRONG></FONT></DIV>
<DIV><FONT size=3D2><STRONG>i think DivX doesn't use "real" =
preprocessing. or=20
the reconstruction filters are very bad,</STRONG></FONT></DIV>
<DIV><FONT size=3D2><STRONG>because it just gives the same =
results than=20
some good cleaners with strong settings.</STRONG></FONT></DIV>
<DIV><STRONG><FONT size=3D2>it seems to be only some washing out =
option (just=20
destroy extremes dct coefs ??)</FONT></STRONG></DIV>
<DIV><FONT size=3D2></FONT> </DIV>
<DIV><FONT face=3DArial size=3D2><STRONG>PS : what's a "symmetric =
7-tap low-pass=20
one" ??</STRONG></FONT></DIV>
<DIV><STRONG><FONT size=3D2></FONT></STRONG> </DIV>
<DIV><FONT size=3D2>MarcFD <A=20
=
href=3D"mailto:marc.fd@libertysurf.fr">marc.fd@libertysurf.fr</A></FONT><=
/DIV></BLOCKQUOTE></BODY></HTML>
------=_NextPart_000_0025_01C23F06.6C002E20--