Re: Differential intra-interchannel filters for RGB images lossless compression


<giorgio.tani@xxxxxxxx> wrote in message
news:1134395838.135883.70530@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
>> if there is a better alternative to png, that manages to gain some
>> popularity (eg: gimp support), I will be interested.
> AFAIK there are projects to use better compression algorithms on PNG, I
> read something about using the 7-Zip's optimized implementation of
> deflate that gives slightly better compression.
> PNG differential filters seem (referring to Waterloo Colorset) to be
> significantly less efficient than JPEG-LS/JPEG2000 (lossless) filters
> on the photographic images, but seem dramathically more efficient on
> computer graphic images (according to literature results in PNG
> documentation and with some tests I just made with the ISA compressor
> linked by Nils, with .jp2, .hck and .isa formats), so, depending on the
> test set, PNG may seem more or less efficient than lossless JPEG
> related...

well, what is likely better depends heavily on how one defines "computer
generated".

"computer generated" as in, eg, images of things like desktops, gui apps,
....
yes, I will agree that png is likely quite efficient.

computer generated, as in, output from a (modern) 3d renderer. I doubt it.
in my experience the composition of these kinds of images is much closer to
that of realistic images.

> Shifting from filter stage to compression stage, as you may note in my
> work, the compression algorithm chosen is very important in determining
> output size, with PAQ giving awesome results at the cost of very high
> computational complexity and with LZMA seeming a reasonable tradeoff
> between compression and speed; however, even with a compressor powerful
> like PAQ a proper filtering of the input data is very important to
> achieve better results than compression on plain, unfiltered data.
> This is consistent with the projects about giving PNG a better
> compressor limited by a tradeoff in speed that on today's machine would
> be acceptable.
>
makes sense.

for realistic images, it is likely that a pure entropy coder would be quite
sensible. for less realistic images, then one capable of dealing with
patterns could make sense.

alternatively, lower emphasis could be placed on the data compressor
(instead opting for something more like a markov predictor for the
off-chance that a repeating pattern appears, and saving in that, for "most"
images, usable repeating patterns are likely a rarity). instead, a faster
coder is used for the filtered data.

insteresting could be, likely, the use of "better" filters.
mental verification of the ones given before is difficult. experimental
testing may be necessary.

the "incomplete" filter:
p=(2*l5-l2)+(2*l7-l6)-(2*l4-(l1+l3-l0))

seems to work ok in a few simple tests (gradiants and edges), but only if
the gradiant or edge is aligned in a certain way:
1 2 3
2 3 4
3 4 p
p=7 according to this

a "complete" variety would be able to better predict a more or less
arbitrary gradiant and likely better deal with edges. right now I am not
quite able to verify that I have it anywhere near correct.

analogy I guess is a kind of table. it can be tilted, and the prediction
moves with the surface. if the table is bent, the opposite corners are
effected, and pulled down to match the tilt.
however, I only have context to the up-left, so there can be problems. if I
could predict, eg, in the middle, then likely more accurate would be
possible.

weird, if likely somewhat impractical, would be to try to reconstruct an
image by regenerating pixels in some possibly bizare order, in order to
allow patterns that more effectively allow predicting values "in the
middle".

or something...


.

Relevant Pages

  • Re: Need help from those here less than 2 years . . .
    ... > don't know why the pages load so slowly, I thought I did everything I ... Might I suggest you export them as PNG rather than JPEG files? ... compression scheme which creates giant files when it comes to photographs, ... but tends to compress images like yours extremely well. ...
    (rec.arts.sf.science)
  • Re: How to specify COMPRESSION level when saving (loseless) PNGs ?
    ... You should not expect to be able to control the compression level in the ... PNG does support a variety of parameters for controlling how the ... But because it throws out data, it works best with images for which you ...
    (microsoft.public.dotnet.languages.csharp)
  • Re: Differential intra-interchannel filters for RGB images lossless compression
    ... AFAIK there are projects to use better compression algorithms on PNG, ... PNG differential filters seem to be ... like PAQ a proper filtering of the input data is very important to ...
    (comp.compression)
  • Re: How to specify COMPRESSION level when saving (loseless) PNGs ?
    ... My PNG are ways too big. ... You should not expect to be able to control the compression level in the way that you are doing. ... But because it throws out data, it works best with images for which you won't notice the loss of data, like photographs or elaborate drawings. ...
    (microsoft.public.dotnet.languages.csharp)
  • Re: New lossless image compression algorithm
    ... white images. ... really correspond to the human eye's three filters, ... The compression software does not try to find correlations between ... pairs of images (eight of the nine cameras on each rover make up four ...
    (comp.compression)