Image Processing Reference
In-Depth Information
Fig. 9.17. The local and the extended models of a facial landmark are obtained by centering
the retina on that landmark in the images of the training set. Adapted from [205]
sponses at each point of the retina. The filters are organised in five frequency chan-
nels and six equally spaced orientation channels as discussed in Sect. 9.7. Filter
wavelengths span the range from 4 to 16 pixels in half-octave intervals.
The same sensor is used both for facial landmark localization and for person
authentication. The sparse nature of the sensor and the relatively low number of
fixations, thanks to the quick convergence of the saccades to the facial landmarks,
make the computation of Gabor responses by direct filtering in the image domain
more advantageous [205] and feasible even in real time [27].
The
saccadic search
strategy requires the use of two levels of modeling for each
of the
facial landmarks
: left eye, right eye, mouth+nose. The latter is a metafeature
that consists in joining the mouth and the nose. The
local model
is obtained for each
facial landmark as the vector of Gabor responses extracted at the central grid point
at the location of that landmark in the images of a training set. The
extended model
is obtained by placing the whole retina at the location of the facial landmark on the
training images (Fig. 9.17) and collecting the set of Gabor filter responses from all
of the retinal points. The local model and the global model are given to supervised
classifiers
7
that learn a compact representation of the facial landmarks by examples,
to identify them later in other images.
Facial landmark detection is achieved by a saccadic search, starting with the
retinotopic sensor placed at random on the image. The search initially aims to find
any of the three facial landmarks. Gabor response vectors are computed at all reti-
7
Examples of supervised classifers are neural networks [29, 49], and Bayes classifiers [125,
153]. and support vector machines.
