Re: Image Compression algorithms

Vision List Digest: Article 7, Volume 10, Issue 19
From: Steve Eddins
Post-Followup: submission@VISLIST.com

Richard Hubert writes:

>Take a look at the JPEG ISO imaging compression standard that will be
>completed by this fall.
...
>It uses frequency domain transformations (Adaptive Discrete Cosing Transform)
>to reach lossless compression levels of at least 20:1 and decimated
>compression of 40:1. If your looking for something simpler, there are lots
>of public domain tools to reach a compression of 2:1 or so.

Curiously enough, I just received a copy of the 8/14/90 draft standard
today. I'm not sure exactly what he means by "decimated compression,"
but I believe the statements regarding lossless compression are
inaccurate on two points:
* DPCM, not ADCT, is used for lossless coding.
* A lossless compression ratio figure of 20:1 is too high.

Here are a few lines from Section 5 (Overview) of the draft standard:

"Two classes of compression algorithms are defined, a set based on the
two-dimensional discrete cosine transform (DCT) and a set based on
spatial prediction techniques. The DCT based algorithms are intended
to give output image quality relative to the source images ranging
from very good to excellent (visually indistinguishable for typical
viewing conditions). The spatial algorithms use differential pulse
code modulation (DPCM) techniques, and are intended primarily for
lossless coding.

...

The compression achieved with these algorithms is dependent on the
characteristics of the images being compressed. For the color images
of natural scenes which have been used for testing and development of
the algorithms, recognizable images are obtained at about 0.15 bits
per pixel and useful images are obtained at about 0.25 bits per pixel.
At 0.75 bits per pixel are typically of very good quality, and at
about 1.5 bits per pixel or less the images are essentially
indistinguishable from the original source images. Lossless coding is
obtained at about 2:1 compression for these test images. All of these
ratios are based on tests of 4:2:2 CCIR 601 format Y,Cb,Cr color
images with an average of 16 bits per pixel."

Steve Eddins
eddins@uicbert.eecs.uic.edu (312) 996-5771 FAX: (312) 413-0024
University of Illinois at Chicago, EECS Dept., M/C 154, 1120 SEO Bldg,
Box 4348, Chicago, IL 60680

http://www.vislist.com