Technology | DOI: 10.1145/1536616.1536623
Kirk L. Kroeker
face Recognition
Breakthrough
By using sparse representation and compressed sensing, researchers
have been able to demonstrate significant improvements in accuracy
over traditional face-recognition techniques.
THe tHeoRies oF sparse repre- sentation and compressed sensing have emerged in re- cent years as powerful meth- ods for efficiently processing
data in unorthodox ways. One of the
areas where these theories are having
a major impact today is in computer
vision. In particular, the theories have
given new life to the field of face recognition, which has seen only incremental increases in accuracy and efficiency
in the past few decades. Now, thanks
to the application of these theories to
classic face-recognition problems, researchers at the University of Illinois
at Urbana-Champaign (UIUC) have
been able to demonstrate significant
improvements in accuracy over traditional techniques.
The idea of applying sparse representation and compressed sensing to
face recognition was so novel in 2007
that two papers outlining the method
were rejected by mainstream vision
conferences. “Just like compressed
sensing, our approach to face recognition is completely unorthodox,” says
Yi Ma, a professor of electrical and
computer engineering at UIUC. “The
reviewers simply did not believe such
good results were possible, or that
sparsity was even relevant.”
Ma had been studying the sparse
representation and compressed sensing theories of Emmanuel Candes and
David Donoho while on sabbatical at
the University of California, Berkeley in
early 2007. It was then, he says, that he
connected the theories to computer vision by applying the ideas to problems
associated with one of the most readily
available sources of raw data for vision
research: face images. “The results were
far beyond what I had ever expected or
imagined from the beginning,” Ma says.
“What happened next was the most exciting period of research I have ever had.”
300
250
200
150
=
100
50
0
0
− 50
100 200 300 400 500 600 700 × +
0
− 10
0
10
20
30
40
50
60
70
80
=
100 200 300 400 500 600 700 × +
The face-recognition method developed at the university of Illinois at urbana-champaign. on the left, the test face is partially covered with
sunglasses (top) and corrupted (bottom). The face is represented in the middle as a sparse linear combination of the training images plus
sparse errors due to occlusion (right top) and corruption (right bottom), with the red coefficients corresponding to the training images of the
correctly identified individual.