Re: WHOA!! Backup the JPG truck...

[P3D John W Roberts <roberts@xxxxxxxxxxxxxxxxx>]

>Date: Sat, 6 Sep 1997 00:09:27 -0500
>From: P3D Jonathan Gross  <catalyst@xxxxxxxxxxxx>
>Subject: WHOA!! Backup the JPG truck...

>For loss-less image compression, seperate out the RGB into three 8-bit
>image channels, and compress each channel using a loss-less LZW (PKZIP)
>method. You'll get between 3:1 and 5:1 compression with no loss of
>important boundary data.

Note, however, that the lossless compression ratio is dependent on the
input data, and the output is sometimes larger than the input (at which
point PKZIP gives up). Lossless compression is therefore not very useful
for cases where the file size / bandwidth must remain within certain limits
(e.g. broadcast HDTV). Also note that using one image and a left-right
difference file (as Norm suggested) has the potential to provide nearly
2:1 reduction on top of the regular image compression algorithm.

It sounds like dividing the original image into the color components
increases the chance of larger regular patterns in the data, which would
help with compression. I'm still surprised by the 3:1 to 5:1 compression
ratio - some people who have made a study of image compression use a rule
of thumb that you can typically hope for about 2:1 compression for generic
images (not counting things like random dot stereograms, which would do
much worse). Has the 3:1 to 5:1 range stated here proven out over a large
and diverse set of images?

John R


------------------------------