lkml.org 
[lkml]   [1998]   [Dec]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: Article: IBM wants to "clean up the license" of Linux
Albert D. Cahalan <acahalan@cs.uml.edu> wrote:
> Many kinds of perceptual compression algorithms for audio and video
> are patented. They remove the data that humans don't notice, and get
> far better compression.

This is an example of a patent I think the free software community
could beat.

(1) Let's imagine that we had some sort of non-linear convolution which
trimmed bits off the image (probably need a couple stages, so it's going
to take a lot of work to design it). So, design a basic framework which
has roughly the right kind of algorithmic properties, and characterize
its parameter space.

(2) Build a big database of before/after compressed images, making sure
that the "after" images are high quality. [Note that you can have more
than one good "after" image.] Even better than a database would be
"too slow but damned good" algorithm that we want to emulate, but I
don't know that there is one.

(3) Run a search in the parameter space of your convolution framework
for a decent transform. Tune it by hand if you can, but just throw
massive parallelism at it (probably make it a genetic algorithm if
you can come up with some kind of "this is good/this is bad" metric)
to pick out a good compression scheme. [Example of how to parallelize:
farm out image pairs to machines, then each machine a stream of parameters
and have them pick the "best n".]

(4) Once you've gotten as far as you can with this system, start getting
real people to look at before/after sets and pick out what they think
is best, and use this as a [Use porn images and you'll have lots of
volunteers.]

Convolution is old hat, non-linear algorithms are old hat, searching and
genetic algorithms are old hat (so are neural net training algorithms,
which I didn't mention but which might be useful here), and polling
people to find out their preferences is old hat. And at no time are
we using a "perceptual model" to design the transform -- we're using
real perception.

This is just a sketch of an idea, and there's some serious room for
improvement, but the underlying concept should work for any lossy
compression algorithm.

--
Raul

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2005-03-22 13:46    [W:0.096 / U:0.312 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site