3. Feature extraction
Following extraction of the lip contour,we create a lip descriptor. For
this we consider two different geometrical–sequential features on two
grid layouts; rectangular and polar. This step transforms the 2D contour
into a one-dimensional feature vector. The rectangular grid features are
the Euclidean distances from sample points along the vertical and horizontal
axes to points on the lip contour. For the vertical axis,we equally
sample 180 points, while for the horizontal axis we sample 300 points
(Fig. 4a). Thus, the resulting normalized feature vector has 480 elements
after concatenation.
The polar grid features aremotivated by thework of [29]. These have
been proven to be a powerful descriptor for offline signature verification.
The set of polar features is calculated from the centroid of the lip
contour. The orientation is sampled from 0 to 360° with a sampling interval
of one degree. For each angle, the radius is computed as the distance
between the point of intersection on the contour and the center
(Fig. 4b). Hence, the resulting normalized feature vector consists of
360 elements after concatenation. We have found that this feature
size gives better accuracy rates.
4. Methods: classification systems and kernel
In this section, classification approaches are described.We used two
classification approaches Hidden Markov Model (HMM) [37] and support
vector machine (SVM) [38]. HMM has been deployed in order to
model the sequential information fromour features, in a similar fashion
to [37]. Finally, these features have been transformed based on anHMM
kernel learnt by SVM.
4.1. Hidden Markov Model
HMMs have becomeincreasingly popular over the past two decades.
They are theoretically sound and usually performverywell in real applications
such as speech recognition [39]. In this section we review the
theoretical aspects relating to our work. An HMM has two associated
stochastic processes; dynamics and observations. The former is not
visible and is usually modeled by the probability of transition between
hidden states. The observation process is modeled by the probability
of obtaining an observed value given a hidden state (see [37,40] for a
complete treatment of HMM). In our paper, we use a discrete HMM
(DHMM) [37] since it forms the basis of ourDHMMKin the next section.
A DHMM consists of the following parameters:
1) The number of states N
2) The number of different observations M