RedBot zaps red-eye in a snap

by Anne Stuart

In the beginning, there were cameras. Then came flashbulbs. Then came ever-smaller cameras with attached or built-in electronic flashes. And with them came red-eye -- that eerie crimson glow in a subject’s eyes. It can make even cutest kid can look downright diabolical.

Red-eye’s cause is simple: The flash bounces back off the blood vessels in the subject’s retina; the camera captures both the burst of light and the reddish hue. Today’s compact cameras exacerbate the problem because of the proximity of flash unit and lens.

"The closer they are, the more red-eye you’re going to get," says Robert Ulichney, a Cambridge, Mass.-based scientist in HP’s Image Systems Lab and one of the HP Labs researchers working to get the red out.

Attacking red-eye on two fronts

HP researchers have developed two approaches to detecting and correcting red-eye. The first exploits an earlier algorithm that accurately detects faces in photographs or real-time video. The second is based on a detector trained specifically to find red eyes.

A Labs-based implementation of these algorithms led to HP’s In-Camera Red-Eye Removal, which instantly removes red-eye from photos while they’re still in the camera, without using a PC and graphics software -- an industry first.

Introduced in May in HP's new Photosmart R707 digital camera, the red-eye removal feature is one of several imaging technologies that originated in HP Labs and are now appearing in a line of new HP cameras.

One-click system

Meanwhile, Ulichney and collaborators J.M. Van Thong, Matthew Gaubatz, and Huitao Luo, have developed a simple, one-click system they call RedBot -- available free online -- aimed at making HP’s imaging algorithms even better.

Beginning in mid-2003, the scientists tested RedBot on about 1,700 photographs submitted by HP employees through a company intranet. The algorithms found and fixed red-eye about 90 percent of the time. "That’s a very good success rate," Ulichney says, but researchers want to push the percentages even higher by running their algorithms on a much larger sample.

RedBot actually employs three algorithms: the original face detector-based one, the eye detector-based one, and one they acquired from FotoNation Inc. of San Francisco. Each photo is tested randomly using one of these algorithms. The Web service logs the details and records feedback from the user.

Improving the algorithms

The RedBot site, which went public earlier this year, lets anyone submit a digital picture for free automatic red-eye correction, a process that usually takes less than one second. In exchange, RedBot retains a copy of the image for use in evaluating how well the process worked.

"The number one goal is to improve the algorithms, and the only way to make it better is to test it," Ulichney says. Specifically, the team wants to gauge how well each algorithm detects and fixes red-eye. Researchers are also looking for instances where the algorithm misses those crimson pupils or registers "false positives," incorrectly detecting something like a reddish facial blemish or a Christmas tree light.

An image that is not corrected is a useful research tool, Ulichney explains. "It’s valuable when we come across something that causes the algorithm to fail because it uncovers shortcomings that may not be exposed otherwise."

Submitting images

IThe team hopes RedBot will harvest at least 10,000 images. To work with RedBot, photos must be in JPEG format, in files smaller than 3.5 megabytes.

Skip the pictures of dogs, cats, and other pets; currently, the algorithms are tuned for human eyes only, Ulichney says. Photo flash reflections in animal eyes are not even red –blue, green, and yellow hues are common. (All photos collected on RedBot are tested privately and won’t be publicly displayed.)

Ultimately, HP’s solutions will provide photographers with an easy alternative to existing red-eye correction methods which are often ineffective or too difficult for most people to use successfully.

Cameras with "pre-flash" capability -- those that flash once to contract the subject’s pupils and again to capture the image --are problematic because most subjects think they’re done after seeing the initial light. "Then, when the actual flash comes, the pose is lost," Ulichney says. Other times, the first, bright flash causes the subject’s eyes to close. In addition, a flash consumes a significant amount of power, so eliminating the need for the double flash extends battery life.

Extending red-eye removal

Fixing red-eye after photos are downloaded to a PC isn’t an ideal solution, either. Most photographers aren’t proficient enough with the graphic arts software to get natural-looking results.

Ultimately, HP will embed software for red-eye correction technology in all its imaging products, an initiative that will provide significant competitive advantage in a rapidly growing market. It will also include the red-eye correction software in its common user interface for all its products.

By 2008, more than a billion small (and thus red-eye prone) -- devices, including cell phones and handheld computers, will take flash photographs, Ulichney notes. If the team has their way, red-eye will be history by then.