Future of the Image Compression

[Nimo]

Hi..,

well, I'm amazing
how would be the future of Image compression in
the coming years..,

If you folks have any ideas( if it sounds weird, no one cares )
please, just shout what you think :_)

starting with me..,

" I'm expecting efficient *pattern recognition* algos
which are going to be included in the respective image
decoders. "

so long
nimo
______
"Je n'avais pas besoin de cette hypothèse-là",

[Industrial One]

http://www.youtube.com/watch?v=lEVe7vU5WiU

Thread: http://groups.google.ca/group/comp.compression/browse_thread/thread/ba0848fd0f3fcba4?hl=en#

[Thomas Richter]

Nimo wrote:
> Hi..,
>
> well, I'm amazing
> how would be the future of Image compression in
> the coming years..,
>
> If you folks have any ideas( if it sounds weird, no one cares )
> please, just shout what you think :_)

Most activity is in the video business by now since you can achieve
higher compression rates there. But anyhow, there are a couple of
"content-adaptive" transformations out there that are being worked on.
Curvelets, Wedgelets, X-lets... Plus approaches that are very specific
for the application domain. Pattern recognition is "too broad" a field,
but the underlying idea of segmenting an image into objects might be.
Using inpainting for compression might be another. What's also currently
under research are "perceptive" coding mechanisms, i.e. how can I
improve the visual quality (vs. the "PSNR") of an image, using here
filter approaches motivated from human perception.

So long,
Thomas

[Nimo]

Thomas Richter <th...@math.tu-berlin.de> wrote in
news:h13858$jkg$1...@infosun2.rus.uni-stuttgart.de: On Sun 14 Jun 2009
09:28:18a, you wrote in comp.compression:

> Nimo wrote:
>> Hi..,
>>
>> well, I'm amazing
>> how would be the future of Image compression in
>> the coming years..,
>>
>> If you folks have any ideas( if it sounds weird, no one cares )
>> please, just shout what you think :_)
>
> Most activity is in the video business by now since you can achieve
> higher compression rates there. But anyhow, there are a couple of
> "content-adaptive" transformations out there that are being worked on.
> Curvelets, Wedgelets, X-lets... Plus approaches that are very specific

I know only wavelets & curvelets; may be I've to update my grey
matter with these latest terms Wedgelets, X-lets.

> for the application domain. Pattern recognition is "too broad" a
> field,

uh.., but there's lot of hype in the OS newsgroups/forums, might be they
are just using that same word for the last 3-5 years without caring
anything about it.

> but the underlying idea of segmenting an image into objects might be.
> Using inpainting for compression might be another. What's also
> currently under research are "perceptive" coding mechanisms, i.e. how
> can I improve the visual quality (vs. the "PSNR") of an image, using
> here filter approaches motivated from human perception.
>

Filter approach.., wow
this would be very intertesting;

working on Lossy instead of *lossles* would be very productive, that's
what I discovered during this course and soon I dropped my efforts
working on it(Lossless); Jules,I know you are listening this :-)

> So long,
> Thomas

Thanks Thomas :-)

so long
nimo

Current status:- working on wavelets

[Thomas Richter]

Nimo wrote:

> working on Lossy instead of *lossles* would be very productive, that's
> what I discovered during this course and soon I dropped my efforts
> working on it(Lossless); Jules,I know you are listening this :-)

However, while there is definitely nothing wrong with lossy compression,
you *must* have some control on the error, otherwise it's pretty
useless. That is "rate distortion theory" is here the main issue, and
similar to the "counting argument" for lossless compression, the
rate-distortion curve is the limiting curve for lossy compression.

To be specific: You *must* define the error measure for which you are
optimal, i.e. cause minimal loss for a given rate.

I.e. while for lossless the uncompressible source is the uniform i.i.d.
zero order source, it is a zero order i.i.d. *Gaussian* source for which
lossy compression works least effective - with the mean square error as
error measure, that is. What is happening for other distortion measures
is not yet known, so there is work to be done. The mathematics for that
is not so easy.

So long,
Thomas