Using the detected skin contour and some simple heuristics, a window
can be defined which is expected to contain the eyes. We then propose
the use of the dark symmetry transform [3] [7] [9]
[6]. This is an annular sampling region which detects edge
configurations that enclose an object. However, unlike template
matching, a perceptual measure of symmetric enclosure is computed and
blob centers are detected. When applied at the appropriate scale
within a window defined by the skin contour, this transform
consistently detects the eyes in the face. The dark symmetry transform
is computed from a phase and edge map by wave propagation (for
computational efficiency). For each point in the image p, at each
scale or radius r and for each symmetry orientation
we find
the set of cocircular pairs of edges
.
The
magnitude of axial symmetry in the (p, r,
)
space is as
follows:
where
and
are the edge intensities of the two co-circular edges and
is the angle separating their normals.
Then, radial symmetry is determined from the axial symmetry map as in Equation 3 and Equation 4. Finally, the symmetry map undergoes Gaussian smoothing and local maxima are determined.
The strongest peaks of dark symmetry are candidates for eye positions. Simple heuristics are used to reject pairs of eyes that have insufficient intra-occular distance (w.r.t the skin blob) and that form an angle larger than 20 degrees from the horizontal. The interest map resulting from the dark symmetry transform is shown in Figure 4.
Horizontal limb extraction is performed to find the mouth from the dark axial symmetry map. The longest linked limb is selected as the mouth.
Additionally, a coarse estimate for the nose's vertical location is found by searching for the strongest vertical gradient in the intensity image that lies in a region bracketed by the eyes and the mouth.
At this stage, a variety of candidates have been detected as possible facial features. These candidates must be tested by more discriminating techniques to discard false alarms and to refine localization.