When less is more: New image compression success could save millions

We all know the pain of sending a too-large photo via email, only to have it bounce back into our can’t-send file.

Or how about on the other end -- the agony of pasting a too-small photo into a too-big web hole, and ending up with a fun-house pixelated version of the image.  The photo was so compressed, too much of its visual information disappeared.

Now imagine you are NASA, trying to capture high-resolution images of the Martian landscape and then transmit them across millions of miles of space over limited bandwidth channels – losing as little information as possible while achieving as much compression as possible.  Even with our narrow experience doing Earth-bound photo transmission, we can tell this is complex and really hard.

If you are Frank Moore, a computer science professor at UAA, this is your problem to solve, which you happen to be quite excited about.

That’s because Moore has had recent success “squeezing” images through compression, but retaining high resolution, results that promise to go beyond the best NASA has ever seen.

First, a brief lesson in compression. To compress an image, you begin by substituting a coefficient for every pixel. Nothing is compressed at this stage, just transformed. Next these coefficients can be compressed in waves so that the information is statistically concentrated in fewer and fewer coefficients. The result is smaller files that are easier and cheaper to store or transmit.

But, there is a trade off. On the other end, when you receive the image, it needs to be reconstructed as accurately as possible. Somewhere, there’s a sweet spot between high-resolution images and file size, and Frank Moore has found it.

State-of-the-art methods use “wavelet transforms” for compressing and reconstructing images. NASA created its own family of wavelets called ICER.  They use “lossy” compression (that’s right, you intentionally lose some of the visual information in exchange for a smaller file).

Notice the pixelation in this ICER transformed Mars detail.

Moore’s work has shown that you can use an evolutionary algorithm to optimize new sets of numbers that correspond to new transforms capable of outperforming these contemporary wavelets.

He started out by showing that --  for an equal amount of compression --  you can significantly reduce the error when reconstructing the image. Then, he showed that -- accepting a given amount of error -- you can reduce the compressed file size further.

His most recent work has, for the first time (drum roll here...), reduced both the compressed file size AND the error rate. Better still, file size and error rates only improve with additional levels of compression. "That's not something that happened with earlier results," Moore said.

He has numbers to demonstrate how his error rate and file size evolve with added levels of compression.

• The best single-level compression yields a 26.8 percent reduction in resolution error, but only a 3.3 percent reduction in file size.
• At three levels of compression, the best-evolved transform yields a 45.9 percent error reduction, with file sizes 13 percent smaller.
• But the big success came with five levels of resolution. Here, errors were reduced by 50 percent and the file size was 28 percent smaller.

Moore's version shows less error in image reconstruction.

The worldly applications for this compression and reconstruction success are huge. Smaller files that still effectively reconstruct an accurate image mean cheaper storage and transmission costs, something government and industry need and want.

Medical imaging is a good example. The price of storage and transmission has been billowing at an annual rate of 50 percent; billions could be saved with cheaper but still accurate compression methods.

Other applications, for us mere Earthlings? Moore says his new transforms could produce higher-quality mp3s, instead of the hollow-sounding versions we can now cost-effectively create.

He’ll use his recently awarded INNOVATE funds to further his evolutionary computations, applying his NASA success to an international compression standard.

Read more about his work in his own words; five projects are posted on his faculty website.
__________________________________________________________________________

Frank Moore with "The Great One."

Frank Moore is an Associate Professor of Computer Science at the University of Alaska Anchorage. He has taught computer science, computer engineering and electrical engineering courses since 1997, and has more than six years of industry experience developing software for a wide variety of military research and development projects.

His current research uses evolutionary computation to optimize transforms for lossy image compression and reconstruction, funded by a NASA EPSCoR CAN for Research award. Moore has published more than 75 peer-reviewed journal articles, conference papers and technical reports.