RE: Data compression

["andreas axelsson" <andreas.axelsson@dynamics.saab.se>]

Hi,

I just took a quick look at:
g0483977     TASS I-band image     800x896   1404KB
g0493955     TASS I-band image     800x896   1404KB

with WinZip at highest compression I achieved a reduction of 36% down to
904KB resp 900KB.

Then I calculated the information entropy in matlab of the images without
any model at all
and that gave 8.94 bits and 8.27 bits per pixel thus a reduction of 44% and
48% down to 785KB
resp 725KB.

In other words, with pure entropy coding, knowing only that each symbol was
16 bits, I reduced the
size losslessly with 44-48% in this case. Having some kind of model too
would give better yield.
The entropy coder would probably be an arithmetic coder since it's more
efficient and
optimal than huffman.

What level of compression do you think is necessary for lossless before it's
worth the effort?
It would be nice to know when working on modelling.

Lossy coding is another matter. I think a vector quantizer before the
entropy coder would
give better compression relative compression error than just removing the
lowest bits.
Especially as I understand you don't want to reduce the quality so much that
any future
novel analysis of the data might be jeopardised.

best regards,
Andreas