[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[infowar.de] Getting a Pixel Fix on the Enemy
Getting a Pixel Fix on the Enemy
By Erik Baard
2:00 a.m. July 17, 2002 PDT
Our culture thrives on simultaneously preserving data over eons and
cramming it through a pipeline in a blink. Now two mathematicians
working for the U.S. Navy have developed an algorithm that promises to
do a bit with both, aiding experts who need to see the world through the
eyes of old artistic masters and killer robots alike.
The Office of Naval Research is funding the work of Guillermo Sapiro,
of the University of Minnesota, and Andrea Bertozzi, of Duke University.
The Navy is motivated to dispel the "fog of war" -- making sure visual
intelligence is complete and in a form suitable to the human eye, as
well as available in real-time.
Transmission troubles can corrupt the data of digital images and
video, leaving images blurry or blotchy. Compression to save bandwidth
can also degrade images. Sapiro and Bertozzi, an expert in fluid
dynamics, recover lost information in its most probable form by
extrapolating on the color, shading, lines, and other qualities of the
region surrounding the gap.
"The algorithm is automatic. It looks at the gray values and gradients
(edges) of the pixels surrounding the hole," Sapiro said in an e-mail.
"In simpler words, both the color/intensity and the borders. These
borders are continued inside the hole."
Art restorers often deal with images made incomplete by stains, rips,
scratches or flaking paint. These experts take days or even months to
"inpaint" a bridge across the gaps to create a consistent image, using
knowledge and subjective intuition in what is an ancient craft.
In the world of mathematics, that's called "partial differential
equation-based interpolation of lost image regions," and it's done at
lightning speed and utterly objectively. By contrast, a software program
like PhotoShop digitizes the process but doesn't change the essentials
For the Navy, apart from correcting faulty images, this new form of
inpainting could dramatically cut back bandwidth demands in the field by
having crude blocks of basic information sent from a source, with a
machine on the receiving end sharpening and filling the image with the
algorithm. In what's sure to get them e-mailed images of UFOs, grassy
knolls and Nessie, the researchers say the technique might be applied to
But some of Sapiro's image alterations leave the viewer wanting to
reach for his or her Photoshop stylus. The larger a filled-in gap is,
the smudgier it seems to be. The program ably removes text from a
distant image, revealing a clean photo underneath, because details are
hard to discern anyway. But close-ups of larger obstructions or gaps
leave a ghostly blur, almost as if the neighboring pixels are
watercolors bleeding into the void.
"In many cases (we) perform perfect, in many not. We follow basic
rules used by restorators. But in (a) half-minute/minute, we do not get
always what restorators do in days/months. Sometimes, the restorator has
to briefly go over our results, and use them just as a hint or as a
starting point. "It can be a fully automatic tool for some cases and a
helping tool in others," Sapiro said. The program sees only pixels ?- it
still takes organic intelligence to discern an artist's intent, or to
mimic style like the scintillation of a Klimpt or the folds of fabric in
"This is a really interesting question," Bertozzi said. "My group at
Duke is just starting to explore, on a more rigorous mathematical level,
how to quantify this more precisely. The methods we are using are based
on very nonlinear processes for which such questions are an active area
of research in ... the physics community."
But there can still remain a sore need for human judgment -? in one
image a bungee cord is removed so the subject appears to be flying over
river. Sadly, the jumper still has a pronounced wedgy lifted toward the
"We never claimed, not in our publications and not in our talks, that
we perform 100-percent perfect," Sapiro concedes.
Mail an infowar -
- infopeace -
de mit "unsubscribe" im Text.