Next: Selective Symmetry Detection for
Up: RealTime Symmetry Transform
Previous: Dark and Bright Symmetry
An algorithm very similar to Sela's proposed attentional mechanism is utilized
to produce interest maps. These interest maps peak at the intersections of
lines of symmetry which occur at the loci of symmetrically enclosed regions or
``blobs''. Certain modifications to Sela's implementation were made. These
include the ability to change the range of the values of r of the annular
sampling regions, as well as the ability to select between dark, bright and
general symmetry. Furthermore, the symmetry lines are kept as an output of the
algorithm since they will be utilized to identify ``limbs''. Limbs are
symmetric structures with a single salient line of symmetry. Since this single
axis is not intersected by other lines of symmetry, limbs do not generally
trigger a strong response in the interest map. However, Kelly [20]
claims that limbs can have significant perceptual significance despite this.
Consequently, the intermediate data are very useful in the extraction of elongated
limblike structures, as proposed by Kelly. Limb extraction will be illustrated
in the Chapter 3 as a technique for extracting the mouth from a face.
The following illustrates a typical application of the algorithm to compute
the general symmetry transform of an image. Figure displays
the input to the algorithm. The lines of symmetry are computed over all 8
scales (i.e. at all 8 values of r) using Equation .
Figure shows the resulting line segments or points of symmetry
which must be linked to form continuous lines of symmetry. The scales used for
these symmetry maps are r=1, r=2 and r=8. Figure shows
the effect of Equation which combines the 8 separate
maps by selecting the maximum response from r = 1 to 8. The desired
interest map I(p) is also shown in Figure . The points in I(p)
undergo Gaussian smoothing and local maximum detection [42] to generate
a set of discrete interest points at the centers of clusters of response found
in I(p). The resulting interest points are finally displayed superimposed upon
the input image.
Figure 2.10:
Input to the attentional
mechanism. (a) Original intensity image. (b) Non maximally
suppressed, thresholded, Sobel edge map. (c) Phase map.

Figure 2.11:
Lines of general
symmetry at multiple scales. (a) General symmetry lines at r=1. (b)
General symmetry lines at r=2. (c) General symmetry lines at r=8.

Figure 2.12:
Combining lines of symmetry
(a) Maximum symmetry lines over all scales. (b) Resulting
interest map, I(p). (c) Smoothed peaks of interest map superimposed as + signs on the original image.

Next: Selective Symmetry Detection for
Up: RealTime Symmetry Transform
Previous: Dark and Bright Symmetry
Tony Jebara
20000623