changes in illumination, resolution loss and variation in facial
expression.
• The DHMMK requires fewer training images per subject.
• The rectangular grid featurewas found to provide better performance
than the polar grid feature.
• The proposed approach is less sensitive to HMM system parameters,
such as the number of hidden states, N.
• The proposed approach is found less sensitive to scale and resolution
changes.
Previously, using static lip features with an HMM has not been particularly
promisingwith regard to subject authentication [14].However,
DHMMK overturns this perception and shows great promise for biometric
based access control. This is due to the fact that the DHMMK captures
valuable information from the gradient of the probability of
having a certain feature vector in a particular state for both the rectangular
and polar grid features. Furthermore, the fact that it works well
while only requiring two training images is important with respect to
subject registration — an important practical consideration for large
scale access control.
Our lip extraction technique is robust to the change of skin colors to
some extent. This was demonstrated by the lip contour extraction results
on the PIE dataset illustrated in Fig. 3. The first image presents
much darker skin color than the second image. The lip contour could
be extracted relatively reliable in such cases. Much darker region such
as a black face might affect the lip extraction results heavier. One possible
idea to overcome this is to develop a possible face detector and adapt
the lip extractor for black, dark, and Caucasian faces respectively.
It is worth pointing out that in all the three datasets tested, it is not
necessary to conduct rotation correction since there are no rotated lips
presented. To handle the moderate rotations of the lips, one possible
way is to develop a simple algorithm in the pre-processing block to