Path: utzoo!utgpu!watmath!iuvax!ndcheg!uceng!mfinegan
From: mfinegan@uceng.UC.EDU (michael k finegan)
Newsgroups: comp.graphics
Subject: Re: SigGraph Fractal Compression
Message-ID: <1934@uceng.UC.EDU>
Date: 17 Aug 89 15:55:24 GMT
References: <444@mit-amt.MEDIA.MIT.EDU> <20400001@inmet>
Organization: Univ. of Cincinnati, College of Engg.
Lines: 36

rich@inmet writes:

>I suppose people are referring to the Michael Barnsley's demo at the AT&T Pixel
>Machine booth.
~
~
~
>Remember decompressing really means applying fractal equations over and over 
>again (like in painting a Mandelbrot).  The amazing thing is that he was 
>"decompressing" the pictures at video rate: 22 pics per second.  

    I have seen the Pixel Machine ads: "up to 820 MFLOPS ...". Is the fact that
he could decompress quickly on a Pixel Machine meaningful? Wouldn't you have to
know how many DSP32's (~10 MFLOP) were being used (i.e. how many MFLOPS) and
then talk about how many MFLOPS/(row X col picture) to decompress? At 820
MFLOPS/sec the demo you saw could have been 37 MFLOP/picture to decompress
(size of images?), which might be a higher degree of complexity than for
discrete cosine transform, etc.

    For comparison, the AT&T DSP Review (Winter '89) shows an example of the
use of a DSP16A (integer based ~ 30 MIPs) to perform DCT of full-color (24
bits/pixel) images, with compression of 32:1. That would transform a 3/4
Meg input file (512 X 512 X 24bpp) into ~24K output file. The transformations
(compress or decompress) each take 1/2 second (at ~30 MIPS).

    While MIPS and MFLOPS are different animals, the DSP16's could also be
used in parallel, and at 820 MIPs the DCT methodology would yield ~54
pics per second (compress or uncompress). This makes the fractal decompression
look pretty good (1/2 performance of current methods); what about fractal
compression? I was under the impression that encoding the image using 'fractal
analysis' took many orders of magnitude longer than the decoding ...

				      -	Mike Finegan

					finegan@aicv01.ece.UC.EDU
					mfinegan@uceng.UC.EDU