identification techniques based on images have received considerable
attention recently, but little attention has been paid to information
in the image of the ear. This work demonstrates the effectiveness
of detailed analysis of ear images, using the NIST database of
real mugshots. Two innovative methods are aplied to boundary analysis.
First, edge analysis is performed only along rays emanating from
a point near the center of the ear, with time and quality advantages
over traditional methods. Second, the innovative concept of "interpretation
breeding" is introduced: two contrasting methods for finding
the ear boundary are combined. Ear images are then cut out and
standardized in several ways to compensate for image variations.
For identification, a neural network is used to compute a distance
criterion derived from several criteria, including components
of an "eigenear" basis similar to Pentland's eigenfaces,
one based on comparison of the most robust portion of the boundary
curve, and another using an eigenbasis of relevant subregions.
The best match to a random query is found 58% of the time, and
the correct match is among the top 5 77% of the time. These results
compare favorably with those for frontal images from the NIST