Abstract:
Although many system
s
exist for autom
a
tic
classification of faces according to
their em
otional expression, these system
s
do not explicitly estim
ate the strength
of given expressions. In th
is thes
is a
n
algor
ithm
capable
of
e
s
tim
ating th
e degre
e
to which a face expres
ses a given emotion
is des
c
ribed and empirically evaluated.
The system first align
s
and norm
a
l
i
zes an inpu
t face im
age. It then applies a
f
ilter bank of
Gabor wavelets and re
duces the data’s dim
e
nsionality via princip
a
l
com
ponents
analysis (PCA). Finally,
an unsupervised F
u
zzy-C-Mean (FCM)
clustering algorithm
is employed to induce
a set of cluster m
e
mberships, which
are then m
a
pped to sub-groups (degrees) of
a facial expression (i.e. Less Happy
(LH), Moderately Happy (MH), and Very Happy (VH)). This unsupervised
m
e
thod is used to determ
ine the best
pair of
Principle Com
ponents and the
centroieds of the clusters for latter clas
sification. To test the hypothesis facial
express
i
on from
the Carnegie M
e
lo
n University (CMU) facial exp
r
ess
i
on data
base is used. The test re
sults on four basic em
oti
on’s (Happy, Surprised, Angry
and Sad) degree estim
ation reflect the
hypothesis. The accuracy is m
easured
em
pirically which shows it is poss
i
b
l
e to es
tim
a
te the degree of
facial ex
pression
using the FC
M algorithm.