Feature Detection and Outline Registration in Dorsal Fin Images

1
Feature Detection and Outline Registration in
Dorsal Fin Images
A. S. Russell, K. R. Debure, Eckerd College, St.
Petersburg, FL
Abstract
Outline Registration
Marine mammologists studying the behavior and
ecology of wild dolphins often employ
photo-identification as a means of associating
observational data with individual dolphins.
DARWIN is a computer program that addresses the
difficulties of manual photo-identification by
applying computer vision and signal processing
techniques to automate much of the process. The
DARWIN system allows a researcher to query a
database of previously identified dolphin dorsal
fin images with an image of an unidentified
dolphin's fin. The researcher may then browse a
rank ordered list of database fin images that
most closely resemble the query image to identify
the dorsal fin. Since the examination begins
with fins most similar to the unidentified fin,
the time required of the researcher to identify
the correct match is potentially reduced. A
major challenge in the automated process arises
from the presence of perspective distortions,
which can cause different views of the same
dorsal fin to differ significantly in appearance,
making direct comparison extremely problematic.
Current research has focused on methods of
distortion correction to quickly transform images
so that it appears fins were photographed from
the same angle. Locating salient fin feature
points provides a basis for this transformation
by allowing a direct determination of any
rotational, translational, and scaling factors
necessary to compensate for perspective
distortions. Preliminary results show that this
approach produces appropriate transformations
when the feature points are accurately
identified. The resulting distortion correction
produces a marked improvement in the similarity
of the dorsal fin outlines. This new approach
promises improved accuracy and a compelling
alternative to manual photo-identification.

• Ideally, we would like to transform the outlines
  so that it appears the original fins were
  photographed from the same angle. The
  transformation process is as follows
• locate sets of common feature points on pairs
  of fin outlines
• compute transformation matrix which maps one
  set of points onto the other
• apply transformation to entire outline

Direct comparison of chain code representations
is problematic since perspective distortions can
cause outlines extracted from multiple
photographs of the same dolphins dorsal fin to
differ significantly. We present a technique to
transform the outlines so that it appears the
original fins were photographed from the same
angle. Locating common feature points on fin
outlines provides a basis for this transformation.
Automatic Feature Selection

• Starting Point of the Leading Edge
• compute absolute angles between successive
  points along the edge
• identify threshold angle which maximizes
  between class variance 3 of the
• angles which comprise the proper leading edge
  of the fin and the angles
• which are associated with the body of the
  dolphin.
• discard line segments at leftmost end of
  outline if their angles diverge
• significantly from predominant edge
  orientation
• select the initial point of the leading edge
  if no such divergence exists

Introduction

• Tip of the Dorsal Fin
• compute a wavelet decomposition
• of the chain of angles comprising
• the fin outline
• identify largest positive local
• maximum in a coarse level
• representation of the outline
• (roughly indicates the position of
• the fin tip)
• track position back through the
• finer scale representations to more
• accurately identify the position.

Figure 1 Following a query of the dorsal fin
database, DARWIN presents icons representing fin
images which most closely resemble the unknown
fin image.
Dorsal Fin Identification

• Most Prominent Notch
• analyze wavelet decomposition of angles
  comprising the dorsal fin outline,
• localizing search to the trailing edge of the
  dorsal fin
• identify candidate notches as local minima
  with large magnitudes at an
• intermediate transform level
• track candidate notches to coarser level.
  Identify most prominent notch as the
• minimum that decreases most slowly in
  magnitude
• back-track to the finest level of detail to
  accurately identify position of notch