Slashdot Mirror
Breakthrough In JPEG Compression
Kris_J writes "The makers of the (classic) compression package Stuffit have written a program that can compress JPGs by roughly 30%. This isn't the raw image to JPG compression, this is lossless compression applied to the JPG file. Typical compression rates for JPGs are 2% to -1%. If you read the whitepaper (PDF), they are even proposing a new image format; StuffIt Image Format (SIF). Now I just need someone to write a SIF compressor for my old Kodak DC260."
I would have thought that rather than 'zipping' an existing image format to create a new one just to save 30%, they'd be better off improving the original image compression algorithm or coming up with a new one.
Quite a while ago (years!) I had a program which could compress images into a fractal image format. It was amazing - the files were much smaller than JPEGs but looked a lot better. The only drawback was that it took ages to compress the images. But with the extra CPU horsepower we have today I'm surprised fractal image compression hasn't become more prevailant. It would still probably be useless for digital cameras though as it would probably be impossible to implement the compression in hardware/firmware such that it could compress a 6+ megapixel image within the requisit 1-2 seconds.
Does anyone know what happened to fractal image format files (.fif) and why they never took off?
The linked page shows average decompression times of 6-8 seconds for 600-800 KB files, rising with the size of the file. Who would benefit from this? It's obviously too slow to speed up web pages, and would be far too CPU intensive for consumer cameras. Professional photographers would have no use for this since they would use RAW images.
I mean, it's cool and all to be able to compress JPGs by that much more, but the size gains are negated by the time it takes to decompress them. This seems just like those super high compression algorithms that have rather amazing compression rates, but take -forever- to compress or decompress, making them unusable. The difference is those are obviously and labeled as simply for scientific research into compression, but Aladdin seems to be trying to market this product for public consumption. The listed uses ( http://www.stuffit.com/imagecompression/ ) seem trivial at best.
Who's gonna be buying this?
-Cliff Spradlin
The linked page did not answer some of my questions:
1. Does this only work for JPEG, or also for other (compressed or plain) files?
2a. If it only works for JPEG, why?
2b. If it works for others, how well?
Anybody who can answer these?
Please correct me if I got my facts wrong.
Stuffit repackages its expensive compression software every year (it seems). Now I would be happy to admit my mistake, but their main area of expertise seems to be marketing. Not technology. I reckon JPEG2000 or any number of other newer-than-JPEG formats already exceed whatever Stuffit purports to have accomplished. I suspect they are trying to tie their name to JPEG as a marketing gimmick to win hearts and minds. I doubt this is worth a mention on Slashdot.
We don't need any new standards unless they actually offer an improvement over recent technology. In technology years, JPEG is very old. A replacement for it should do better than 30% and offer other advantages. I suspect a company that actually specializes in imaging might have come up with a better solution a while ago.
Would this technique apply to DV video?
I have a friend who's father is a professional photographer. He has gigabytes and gigabytes of images stored for his customers, should they want to order re-prints. They're thinking about setting up raid terabyte file server. I can certainyl say that this is good news for them!
Electrons are free; it is moving them that becomes expensive.
JPEG is (roughly) a discrete cosine transform, followed by a filter on high frequencies, followed by Huffmann encoding (which is lossless). This is probably the Huffmann encoding that they did remove and replace with one of the more efficient compression algorithms, and something that could indeed be much more efficient than simple huffmann encoding. So they still take advantage of all the strengths of JPEG related to the human perception model, but they still gain in compression. Huffmann is great to compress oft-appearing sequences, and is a great general-purpose lossless encoding, but there are other that do a better job of it.
In other words, you did not understand what they did.
Okay, I do a bit of metalurgical crystalline micrography, and we keep the images in the raw bitmap format because we can't let ANY data be lost. But we also have gobs of drive space, so a good lossless, non-distorting compression format would be handy.
Bacardi + slashdot = negative karma.
The PNG format is lossless, and very widely deployed (pretty much all browsers and image programs, etc).
I rarely criticize things I don't care about.
I thought they had found a method to further compress a JPEG, while still maintaining the original format. i.e. it could still be viewed with a regular JPG viewer. That kind of optimization would've been great, especially if it could be used on webpages, forums etc.
But this is somewhat disappointing. The compression changes the format, and it must be decompressed to view it. Plus they don't intend on releasing the format, and their proposal for a new filetype which can be read by a "plugin" reeks of incompatibility issues.
StrayByte.Net
The quantised DCT coefficients of a JPEG image are compressed using a JPEG standard huffman table. From what I've seen, this table is far from optimized even for "the average of the majority" of images out there.
Ogg Vorbis stores its own huffman table in its own stream. The default encoder uses a table optimized for the general audio you can find out there. There is a utility called "rehuff" (goggle it yourself please) that will calculate and build a huffman table optimized for a particular stream and it seems that on average it reduces an Ogg Vorbis filesize by about 5-10%.
Building an optimized huffman table for individual JPEGs will probably yield such improved compression rates too. If the original JPEG tables are less optimized than the Ogg Vorbis ones, the reduction will be even higher. But 30% seems a little... optimistic.
I wont bother going into the details of how it works (go read 'Fractal Image Compression' by Yuval Fisher*) but I concluded that fractal compression wasnt viable as there wasnt a general solution to suit all images. There are about 20 variables that you can decide, which give variable results to the final compressed image.
And one set of variables would be excellent to compressing pics of trees down 80%, another set of variables would be excellent compressing pics of animals down 80%, but using each others set would only give results equal or worse than normal compression algorithms.
Another factor at the time, five years ago, was that it took an hour to compress a 256x256 greyscale image on a 300MHz machine. Nowadays that isnt a factor.
* If anyone has this on ebook please post a link here, Im dying to read this again.
Okay, so as near as I can figure out, this is a white paper because: .SIF as basically being having everything a jpg does except higher compression.
.sif : bandwidth savings will actually be on the order of 98% (30% image size reduction, 97% website traffic savings).
It describes a new format
Some might think it's a press release, since the "White Paper"'s discussion of the new format is largely limited to the benefits "OEMs" will have in using this (soon-to-be patented) technology, and explaining how it integrates into Allume's fine line of existence and future Digital Lifestyle(tm) products.
Some might further argue that often press releases will accompany white papers, explain the point for the consuming world, and link to the paper; this "white paper" seems to do exactly that: we are invited to examine a (currently 404'd) page to see the actual data involved.
To those who think it's a press release, it's not, because:
It don't say "For Immediate Release"
Press releases are proofread for grammatical errors, and working links capable of withstanding slashdot-class bandwidth
And by the way, they don't have a chance. Sure, bandwidth cost means a lot, but for the OEM who might consider adopting this technology, it means a new standard that offers at best 30% improvement (on those little 50kb jpegs we're asked to believe), paying for it, and passing the cost on to the consumer. It may be an improvement, but if you really want to improve your image-heavy website's bandwidth usage, switch to Fractal Image Format. Not only do you reduce the size of the images; you reduce the number of people downloading them. The same for
I worked in this field for awhile, and the liscencing and other issues sent the company I was with running in the other direction. JPEG was good enough, everyone was using it, so JPEG it was.
Fractal compression is cool.. but encumbered by IP issues. Too bad.
..don't panic
" Typical compression rates for JPGs are 2% to -1%. "
Does -1% mean that the file actually gets bigger?
Big Deal. With broadband ever expanding and storage being relatively cheap and inexpensive, this technology is a waste of time. Hell, I've got a $100 USB drive for X-mas that was a gig...a fucking gig.
If this idea is what's suggested here, ie. instead of
Image -> DCT -> Truncate spectrum -> RLE encode
they use
Image -> DCT -> Truncate spectrum -> some more efficient encode
it wouldn't surprise me if the JPEG Group discussed alternate compression encodings during the standardisation meetings. In fact, I'd be extremely surprised if the matter was not raised. If it was, it's pretty clear that this patent is not novel in the slightest...
Athletic Scholarships to universities make as much sense as academic scholarships to sports teams.
You know what ? The breakthrough in JPEG compression was mainly JPEG-2000.
There had been, was, and has been possible to achieve better compression ratios, and guess what, even in the high times of JPEG we knew that what's inside, it's not the optimal solution. But there were certain aspects which made it stay the way it was, and that was good to be so. The same goes with JPEG-2000. Since the appearance of it there have been many attempts to make it better, and there have been some good results achieved (even I have read and sometimes even reviewed papers dealing with the subject).
It's really no question whether one can make an X% better compression to JPEG with the same quality (expecially today, when JPEG is so old that every joe and his dog had time to develop better ways). The question is, has it enough practical usability to justify its presence ? Is there a well-justified reason why we should use it ? Does it deliver
- better compression rates (smaller size with the same objective & subjective quality) ?
- lower compression times ?
- compatibility ?
- is it any better for hardware implementation purposes (same or lower computation times) ?
If it's just a "better" compression for the compression's own sake and not for our sakes, then this is even less news than me cleaning my room.
I am putting myself to the fullest possible use, which is all I can think that any conscious entity can ever hope to do.
As another person who has implemented JPEG, I vouch for his accuracy :-)
It is this last step which is not particularly efficient. There are several reasons why ZIP cannot significantly compress this data. First, ZIP is a byte-oriented compressor. Huffman codes are bit-oriented. This causes ZIP to miss many patterns that could be backreferenced simply because they fall on a weird bit position. Second, data can be highly compressible and yet have no repetitive patterns in it.
For example, consider the sequence 1,2,3,4,5,6,7,8,9,10,...,1000. It has no repeating pattern, but it would be absurd to claim that it cannot be compressed. In this case, applying the simple transform x[i] --> x[i+1] - x[i] transforms the sequence to all zeros, and that is trivially compressible. To decompress, just apply the inverse transform.
ZIP is too naive to use any sort of transforms. It really was only intended for textual data, program executables, etc.