Archive notice: This article was originally published on January 29, 2008. Links and embedded videos are preserved as part of the historical record.
The algorithms are designed to find similar faces in image databases. “A single photo can be used to train the system,” Hung-Son Le tells pressetext, describing what is special about his development.
The basic handling of the photos is based on the so-called “Hidden Markov Model” (HMM). “HMM is a statistical tool that can be used in many fields to model patterns,” explains Hung-Son Le. He says the method is widely used and already employed for facial recognition. “Previous HMM approaches required many training photos,” the researcher explains, outlining the key difference from his model. It can recognise a face using just one photo as a reference.
“I use another method to improve contrast and visual details,” says Hung-Son Le, identifying an additional strength of his algorithm set. This enables it to handle underexposed or overexposed images and different facial expressions. Sunglasses and, in particular, face coverings nevertheless remain problematic, as they do for all facial-recognition algorithms.
The system was tested using international standards such as the Face Recognition Technology (FERET) database and, according to Hung-Son Le, achieved better results than currently common solutions. Commercial applications of the research findings are already under development, including a web search engine for faces.
According to Hung-Son Le, he has no connection with the likewise Swedish start-up Polar Rose http://www.polarrose.com, whose face search is currently in closed beta. “To my knowledge, it does not work with automatic facial recognition,” he says of Polar Rose’s search. That is also correct, as the company confirms. Although the current beta version can find faces in photos, it does not yet include algorithms for automatic comparison. Corresponding functionality is also being planned.
