Face Localization - Hierarchical Neural Networks for Image Interpretation - page 206

excitatory inhibitory excit. inh. ex. i. ex. i.
Layer 0 Layer 1 Layer 2 Layer 3
Fig. 10.7. Face localization recall. The activities of all feature arrays are shown after ten
iterations when the test image from Fig. 10.3 is presented as input.

The activity pattern forms a sparse distributed multiscale representation of the image
content. The hidden features clearly contribute to the development of stable blobs
at the eye positions and to the suppression of output activity at other candidate lo-
cations, but they are hard to analyze.
The generation of stable blobs is the typical behavior of the network. To evaluate
its performance, one has to estimate eye coordinates from the blobs and to compute
a quality measure by comparison with the given coordinates.
The position of each eye was estimated separately, as illustrated in Figure 10.8.
In a first step, the output unit with the highest activity v max is found in the cor-
responding high resolution output. For all units in a l × l window around it, the
feature cells belonging to the blob were segmented by comparing their activity with
a threshold v t that increases with greater distance d max from the center and with the
activity of the center v max :

v t = 0 . 5 · v max · d max /l.

(10.1)

The weighted mean location of the segmented cells is used as the estimated eye po-
sition. The figure shows the unproblematic segmentation of the test example from
Fig. 10.3 as well as a more problematic case, where a secondary blob has not been
removed by the iterative refinement, but is successfully ignored by the blob segmen-
tation.

(a) (b)

c = .75 c = .82 c = .72 c = .53

c=.61 c=.38
Fig. 10.8. Blob segmentation for face localization. A 7 × 7 segmentation window around the
most active pixel is analyzed. The segmented pixels are framed by a black line: (a) output for
example from Fig. 10.3; (b) output for a problematic example where a secondary blob has
not been removed. This blob is ignored successfully by the segmentation procedure.