News

Three years ago, as he was about to enter the fourth year of an undergraduate biomedical engineering program, a 20-year-old University of Toronto student came up with a novel idea for his final thesis.

Alireza Rabi decided to borrow some of the facial-recognition software created by his professor, Parham Aarabi, the renowned head of UofT's Artificial Perception Laboratory, write some new code of his own, and develop a program that might revolutionize plastic surgery.

The goal was a software program that would use an actual photo to show a person, realistically and in advance, what he or she would look like after having cosmetic surgery. Such a system could put an end to disappointed patients and also help surgeons refine their craft, Mr. Rabi says.

"My father is a plastic surgeon. From his experience with patients I know there is a very great need for tools like this to avoid disappointment with the results of cosmetic surgery," explains Mr. Rabi, now in the first year of a combined MD/PhD degree program at Johns Hopkins University in Baltimore.

"The goal was to create something that would be of equal value to patients and surgeons alike."

The resulting software, dubbed Modiface, allows anyone with access to a digital camera and the Internet to see what he or she would look like with, say, Brad Pitt's eyes or Paris Hilton's nose, or the facial features of anyone else for whom they have a digital photograph.

Still a work in progress, the beta version is now available online at Modiface.com and is proving to be a hit with millions of Internet users. Since January, the site has drawn an average of 100,000 hits a day (when a Toronto television station carried an item about it, one million visitors tried to log in, temporarily crashing the site).

"The early version is intended for consumers, but we are working with plastic surgeons to improve the system and maybe create a commercial application," Mr. Rabi says.

The current version of Modiface is actually the result of 10 years of research into facial-recognition software and about $1-million in research funding, explains Dr. Aarabi, the 30-year-old Canada Research Chair in Internet video, audio and image search.

"What we are trying to do is create a system that first recognizes a photo as a face, and then can easily manipulate the features on that face to create a lifelike and accurate result," the professor says.

Another goal was to do it at great speed. Dr. Aarabi says Modiface can recognize a face within a digital image in just a single second compared with the 10 seconds taken by any other software he is aware of.

"What Alireza was able to do was build on what had been done to date and then create new algorithms to refine some tricky problem areas," says Dr. Aarabi. Chief among those problems was finding a way to blend skin colour and texture when taking a nose, eyes or lips from one face and transplanting them onto another.

"That is the hardest part," says Mr. Rabi. "I had to create all new algorithms so the resulting face looked natural and not like you had stuck someone else's nose on someone's face."

While there are existing systems to simulate the results of plastic surgery, they are generally done by hand and require some training and skill to achieve the desired effect, says Dr. Aarabi. "This one is entirely automated. It is all point-and-click. Anyone with a digital camera and an Internet connection can do it."

The Modiface research is taking place in four distinct stages. The first step was to create algorithms that achieved facial recognition at blazing speed; the second was to create software that would work well in high-traffic areas such as an Internet site. Both those goals have been met, Dr. Aarabi says.

Step three is to have plastic surgeons work with the beta version and provide written feedback. The final stage will be to create a commercial product.

While the ultimate goal is a system for use by plastic surgeons and their patients, both Dr. Aarabi and Mr. Rabi say Modiface will continue to be offered free to consumers through the website, which was created with the help of a $10,000 grant from the National Science and Engineering Research Council.

To date, most facial-recognition research has been into systems that help law enforcement and security forces identify suspects, but results from those efforts have often been disappointing, Dr. Aarabi says. While systems work well in the research lab, they break down in the field.

Two years ago, for example, security staff at the Super Bowl scanned the crowd using a system that had achieved 80-per-cent recognition rates in the lab. At the game, the rate fell to 20 per cent, he says.

"Our goal [with Modiface] was not directed at law enforcement, although the results of our research may have applications there as well," added Dr. Aarabi, who is no stranger to innovative ideas. In 2005, he was named one of the world's 35 most innovative people under age 35 by MIT's Technology Review, for developing a multi-sensor microphone that follows a speaker around a room.

Both he and Mr. Rabi see spin-off benefits already from the current version of Modiface. For example, the program is helping to advance the understanding of artificial intelligence itself.

Dr. Aarabi says the system actually learns from each attempt at reconstructing a face, slowly gaining the ability to correct its own errors (such as putting eyes in the wrong place) and refining its processes. Every three to four weeks the researchers fine tune the system to remove glitches caused by users who might, for example, submit a photo of a dog instead of a person.

"Essentially the software amends and updates its own algorithms automatically as it gains more experience," Dr. Aarabi says.

"By seeing how it deals with learning, we, in turn, get insights into artificial intelligence."

How it works

To use the Modiface site (http://www.modiface.com), users upload digital photos of themselves, or anyone else whose looks they want to adjust. They can then choose from about 30 on-site photos of celebrities (including Brad Pitt, Oprah Winfrey and Paris Hilton) or upload another photo of someone whose specific features they would like to borrow.

"The first thing the system does is recognize the face and specific features," explains Dr. Parham Aarabi of the University of Toronto. "Once that is done, it is a matter of pointing and clicking on the prompts."

The user chooses from a menu that includes hair, eyes, lips or whole face.

Each click automatically superimposes the desired feature onto the subject photo, but there are still some glitches to be smoothed out.

"There are catches sometimes because of shading or light on the uploaded photos," says Dr. Aarabi.

Also, "Modiface can put eyes or lips in inappropriate places," he says. "We have buttons to click to reposition them."

TERRENCE BELFORD