Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!rutgers!network!ucsd!ucbvax!tut.cis.ohio-state.edu!sioux.cis.ohio-state.edu!sarrel From: sarrel@sioux.cis.ohio-state.edu (Marc Sarrel) Newsgroups: comp.graphics Subject: Re: SigGraph Fractal Compression Message-ID: Date: 14 Aug 89 14:56:26 GMT References: <444@mit-amt.MEDIA.MIT.EDU> <20400001@inmet> Sender: news@tut.cis.ohio-state.edu Organization: Ohio State Computer Science Lines: 50 In-reply-to: rich@inmet's message of 12 Aug 89 20:01:00 GMT In article <20400001@inmet> rich@inmet writes: The basic idea is to "compress" an image by finding the probabilistic coefficients of "attractors" equations that when "decompress", gives you back a representation of the image. There are several important points: one is this is not compression in the traditional sense. It is more like modeling. The other one is the decompressed image can be as good as what the output device is. It just takes longer to decompress if the output is larger. Third, compression ratio can be very high. For images with recurring parts (like a picture of a fern), it can be described by 3 equation, each one withe 4 real number coefficient. Thus you need to store 12 32 bits single precision floating point number. Well, this is the best explanation that I've heard so far, but I'm still not convinced of the method's overall usefullness. First, how does one go about finding those coefficients? Is it an automatic process, or does it have to be done "by hand?" How does this method compare to "standard" compression schemes in terms of speed and compression ratio? What is the method's average and worst case image behaviour, you've already given us its best case behaviour? What sorts of images does it work least well on? I could go on, but I have a life to lead... As far as the getting more resolution out than you put in, surely you realize that this is nonsense. You are certainly able to do the calculations, but the added information is just arbitrary and won't necessarily have anything to do with the original information (I don't think). In order to reproduce the image at the original resolution, you only need a certain amount of "accuracy" for your coefficients. (ie: anything after a certain decimal place is noise.) However, it is just this extra information that you're using to "add" information to the image. There ain't no free lunch. (setf soap-box-mode t) I'm sorry, this whole thing smells too much like the "Confusion is a Jar" (tm) debacle that wasted our time a few months ago. I don't trust any scientist or researcher who makes wild claims, but then won't or can't disclose all his reasoning and evidence. (setf soap-box-mode nil) --marc -=- "Master, why is the letter 'i' the symbol for current?" "Because there is no letter 'i' in the word 'current'." "Master, why do we use the letter 'j' for sqrt(-1)?" "Because we use the letter 'i' for current." Whereupon the Master struck the Disciple, and the Disciple became enlightened.