Re: [OT] lossless image compression (was: data compression)

From: Philippe Verdy (verdy_p@wanadoo.fr)
Date: Thu May 05 2005 - 11:53:32 CDT

  • Next message: Jukka K. Korpela: "Re: Cyrillic - accented/acuted vowels"

    ----- Original Message -----
    From: "Elliotte Harold" <elharo@metalab.unc.edu>
    Cc: "Unicode List" <unicode@unicode.org>
    Sent: Thursday, May 05, 2005 2:02 PM
    Subject: Re: Data compression

    > Barring independent verification, I'm skeptical of the claim.

    I would be too, if that searcher had not already called for the service of
    an independant evaluation, and based on his report, obtained funding in the
    Australian stocks exchange.

    > Data compression is a field second only to encryption in the amount of
    > snake oil that's come and gone over the last couple of decades. This is
    > not the first time I've heard spectacular claims made for new compression
    > systems only to never see them actually reach the market. At least with
    > compression it tends to be really obvious that a system doesn't work. With
    > data encryption sometimes the flaws aren't as readily apparent.
    >
    > Perhaps this researcher really has found a technique everybody else
    > missed. But I'll believe it once I see a shipping product. Until then I
    > remain skeptical.

    There's a demo... Really impressive. But this effectively works very well
    for medial imagery, such as radiography, where the images have small
    gradients and for which the traditional lossy compression technics loose too
    much information in those tiny gradient areas, that preserves the smooth
    transitions and precision of the color and light shades.

    I would think at first that the compression technic is an adaptative system,
    which really favors the compression of images with small contrasts such as
    medical clichés (and in some measures the video images, but traditional
    cameras used for cinema and even photography don't have the sufficient
    precision and stability in their CCD cells to obtain stable and smooth
    gradients: see the many unstable pixels which produce additional snowy noise
    on those cameras; if you use lossless compression, you'll need to represent
    and store this unstability, even if the human eye cannot see them; however
    for medical imagery, the captors are much more stable and precise so the
    power of this noise is very low, so this allows capturing the smoothest
    gradients).

    I am not sure it is adaptable to compress any type of image (for example
    bitmaps, computer screenshots, cinema, etc... simply even when they are
    captured with "high definition" cameras; but it may compress quite well the
    computer generated raytraced images which are based on smooth mathematical
    models, provided that these raytraces are not approximated by fast integer
    functions with a too large quantification noise.)

    Note also that medial imagery can obtain very high precision by using very
    long pause time for the capture: the longer the pause time, the more the
    captor will take captured images, that are then smoothed by a simple mean
    function. This increases the precision of the shades, and reduces the power
    of the captured noise (including the quantification noise), but the bad
    thing is that this requires static (immobile) subjects. To reduce the
    capture time, but keep a high precision, mediacal imagery requires a very
    fast sensor capable of taking several hundreds of clichés per second,
    instead of several dozens for cinema, or just a few for numeric photography.

    To summarize: lossy compression is not a bad thing for classic photography
    and video, because it can ignores the inherent noise caused by the capturing
    source itself, and by numeric quantification (i.e. the 2D pixel size
    quantification noise, and the per-pixel energy level quantification noise
    implied by the necessarily limited number of bits to represent it).

    And even for medical imagery, the quantification of the analog signal is
    still required, all captors still have their own inherent noise (caused by
    temperature, or by the effects of the variation of natural environmental
    electromagnetic fields on the captor, or by nuclear desintegration,
    unpredictable high energy particles like neutrinos, gravitation fields, or
    simply the global "white noise" of the universe echoing the bigbang...); add
    to this that mediacl imagery will concern *live* subjects that are not
    completely static during the capture time. Is it still required to keep the
    images with all this inherent noise? May be...



    This archive was generated by hypermail 2.1.5 : Thu May 05 2005 - 11:54:50 CDT