Researchers at Singapore's Laboratories for Information Technology
(LIT) have developed a face recognition technology that can take
into account varied lighting conditions and different facial
positions.
According to Roberto Mariani, a research staffer on the LIT's Media
Engineering Program, the technology is also robust enough to offer
a face recognition system for mobile applications such as robotic
vehicles.
LIT is a national research and development institute controlled by
the Agency for Science and Technology. Showcased at the
CommunicAsia exhibition last week, the technology addresses real
world conditions that seriously affect the accuracy of current face
recognition technology, Mariani said.
Mariani explained that two of these conditions are varied lighting
conditions and varied face poses (the orientation of the face
relative to the camera). The technology has been developed to take
account of varied expressions and the presence or absence of facial
artifacts such as glasses, hairstyles or moustaches. These
conditions can affect either the coding process to create
inaccurate database images, or more significantly, the matching
process when the system encounters a less-than-ideal image to be
matched against the database.
Mariani said the team originally sought to tackle the problem by
exploring how to enhance normalisation techniques to provide good
matches against database images. Conventional face recognition
systems already use various normalisation techniques. Mariani
described this as an "extremely difficult" endeavor, as there were
too many real world lighting and face pose variations for any one
system to adequately accommodate.
The solution the LIT team developed is based on what it describes
as a unique face synthesis procedure. The technology basically
synthesises multiple images from a single image to produce a
variety of images in different poses and lighting conditions.
By boosting the number of realistic images that the face
recognition system can match an image against, the premise is that
at least one of the synthesised images will register a correct
match with the entry in the database. Or, the technology can be
used to increase the number of models in the face database to offer
a richer set of images to be matched against. "The synthesis layer
basically brings more knowledge and intelligence to the face
recognition system," explained Mariani.
With multiple image extrapolation as the foundation of this
technology, the LIT team also incorporated faster matching
algorithms to speed up recognition time, by enabling rapid and
accurate sifting of entries in the image database.
This technology outperformed conventional face recognition systems
in a trial heterogeneous database of 175 images in a wide variety
of lighting conditions, poses and picture quality.
According to the researcher, the solution is also robust enough to
offer a face recognition system for mobile applications such as
robotic vehicles.
LIT is seeking opportunities to work closely with system
integrators and original equipment manufacturers to address key
application areas such as surveillance, access control, mobile
commerce or toy applications.