Re: BTPC Extension - PyramidWorkshop

I hope you're also satisfied with the CALIC+ari results. I may repeat
that with JPEG-LS+ari if you insist. :-)
Well, I think you definitely should test against JPEG-LS here. Because
it is a well-acknowledged algorithm and you will sooner or later hear
questions about your perfomance against JPEG-LS anyhow.

I hope I can present some lossy results in the next time.
Would be useful, definitely.

Okay, in schedule for the afternoon. The UI has a really evil
memory-overwrite bug, so I'm going to switch to the command-line tools.

[...] I assume that JPEG-2000 do not catch inter-planar
properties.
Not really. Inter-plane correlations are only captured, if at all, in
the color transformation. Afterwards, planes are coded independently.
This step was made to make the algorithm more "parallelizable" (is that
a word?), and possibly to fit into the traditional scheme of image
compression, too.

What is that for? Grounding an innovative algorithm on traditional
understanding, only to look traditional?

I think playing strength out is not 'unethic', JPEG-IJG took
subjective advantage of the standard setup YCbCr 4:2:2, which probably
is the strength of JPEG-IJG.
Clearly not 'unethic'. (-; Question is what does it buy?

Quality. In a uncomparable setting 'YCbCr 4:4:4' against 'KLT 4:2:1',
the
latter wins clearly, even with a poor entropy coder. In an equal
setting
the latter is the natural domain for the BTPC algorithm (take
inter-planar
information into account), and should also win clearly.

For those who don't know about KLT: KLT (Karhunen-Loève Transform) is
a
matrix-transform (like YCbCr too) with an algorithmic calculated
matrix,
that (in the 3D-case, means three value-planes) represents an
artificial
color-space in that the biggest turbulences reside in the primary
plane.
Because it has to be a linear solveable system (remember: matrix) you
will not receive absolute sharp distinct planes, rather than having
successive less or unrelated turbulence within the planes after the
first.
The significance for inter-planar information is that (hopefully) most
probably, the most pregnant features in the original color-space are
going
into the primary plane (remember: this is an artificial value-space,
there
not nessesarily exists any valueable 'color' interpretation like in
YCbCr),
and the following planes contain variations.
With that assumption you can inherit the features of the primary plane
to
the other planes (which is (hopefully) most probably right). It is
fairly
possible to drop the other planes completely without receiving a
mono-color
image, possibly it preserves shape and produces 'random'/chaotic
looking
but correlation-based color-propagation into various directions (for
example less green, more blue, ... always after tendencies that occur
in
the original image).
I hope I got a not too wrong non-scientific, non-math, context-free
explanation.

Specifically in the APT, the Zeng-palette-sorting algorithm does a
very similar calculation, ordering the palette after correlation too.
So using KLT as value-space in combination with Zeng is some sort of
"double-adaptive feature reordering" which gives even moderate results
for lossy palettized pictures (for example frymire).

[...] low bit-rate is the domain of vector-quantizers.
Yes. The trellis-quantizer in JPEG2000-2 causes still quite an
improvement, but that's also mainly in the low-bitrate area. JPEG2000-1
uses only scalar quantizers (quite unfortunately, due to patent
restrictions of the trellis/Viterby algorithm).

Have you ever heard a VQ being successfully applied to 0D? I tried to
improve the SQ in various way, but in the end I never find anything
better than dead-zone uniform. I believe, hope, that adaptive
partitioning
of input-space _must_ be better than uniform partitioning, but
everytime
I only end up with more flat, more random, more loss. Even
context-driven
SQs don't work. Semi-adaptive schemes like Loyd would produce too much
side-information (and calculation). Tried to make a neural quantizer
too
(for example 256 input, 12 output), but that one doesn't learn fast
enough
on the run (I don't want trained NN in this case).

So long,
Thomas

Saludos
Niels

.

Relevant Pages

  • Re: Elevators with no floor buttons in the cars -new apps
    ... "elevator algorithm." ... Better than 20 planes from 20 different companies going to the same place at the same time. ... -larry / dallas ...
    (misc.consumers)
  • Re: EM ALGORITHM
    ... I am looking for an EM algorithm that can be used for missing ... Each sensor may be reading a different ... that there can be any number of features and points ...
    (sci.image.processing)
  • Re: EM ALGORITHM
    ... The EM algorithm is known as Expectation Maximization. ... that there can be any number of features and points ... > more with pattern matching and classification. ...
    (sci.image.processing)
  • Re: EM ALGORITHM
    ... The EM algorithm is known as Expectation Maximization. ... that there can be any number of features and points ... > more with pattern matching and classification. ...
    (sci.image.processing)
  • Re: EM ALGORITHM
    ... I am guessing you are looking for the Expectation Maximization Algorithm? ... that there can be any number of features and points ... more with pattern matching and classification. ...
    (sci.image.processing)