there is cancellation between diametrically op-
posed tangent vectors. Axes of symmetry are lines
of local minimum in the tangent ®eld. At the level
of ®ne detail, intensity ridges or ravines give rise to
local symmetry axes.
Unfortunately, since the raw gradient vectors
are likely to be noisy we must develop a means of
smoothing the tangent ®eld so that we can perform
the required topographic analysis. To realise this
goal, we appeal to magneto-statics to develop a
means of smoothing the tangent ®eld. According-
ly, we compute an analogue of the vector-potential
by regarding the edge-tangents as a ®eld of ele-
mentary currents. The vector potential is found by
integrating over volume and weighting the con-
tributing currents according to inverse distance. In
other words, the vector-potential at the point
r x; y; z
T
in the augmented space in which the
original image plane is embedded is
Ax; y; z l
Z
V
0
jx
0
; y
0
; z
0
j
r ÿ r
0
j
dV
0
; 6
where
r
0
x
0
; y
0
; z
0
T
and l is the permeability
constant which we set equal to unity. Since the
contributing edge-tangent vectors (or currents) are
distributed only on the image plane, the volume
integral reduces to an area integral over the image
plane. As a result, the components of the vector-
potential are as follows:
Ax; y; z
ÿ
R R
oG
r
Ix
0
;y
0
oy
0
1
xÿx
0
2
yÿy
0
2
z
2
p
dx
0
dy
0
R R
oG
r
Ix
0
;y
0
ox
0
1
xÿx
0
2
yÿy
0
2
z
2
p
dx
0
dy
0
0
0
B
B
B
B
@
1
C
C
C
C
A
:
7
The structure of the vector-potential deserves
further comment. In the ®rst instance, the com-
ponents are con®ned to the x±y plane for all values
of the auxiliary co-ordinate z. However, as we
move away from the image plane the role of this
auxiliary dimension is to average the generating
currents over an increasingly large area of the
original image plane. In other words, the role of
the auxiliary z-dimension is to allow us to perform
volume integration of the edge-tangent vectors. By
sampling the vector-potential for various x±y
planes at increasing sampling height z above the
image plane, we induce a scale-space representa-
tion. We exploit this property to produce a ®ne-to-
coarse image representation as we sample the
vector-potential at increasing sampling heights
above the physical image plane.
In order to develop the appropriate dierential
operators for feature characterisation from the
vector-potential we have taken the magneto-static
analogy one step further and have appealed to the
geometry of the associated magnetic ®eld. Ac-
cording to magneto-statics, the magnetic ®eld is
the curl of the vector-potential. It is important to
stress that because it is less computationally
tractable than the vector-potential, the magnetic
®eld is never used directly in our image represen-
tation. The role of the magnetic ®eld is to provide
an auxiliary representation. The geometry of the
magnetic ®eld allows us to understand the dier-
ential structure of the vector-potential. According
to our representation of the image structure,
symmetry lines follow the local minima of the
vector-potential. In other words, they connect
image points where there is strong cancellation
edge tangent vectors associated with symmetrically
placed object boundaries. By contrast, edge-
contours follow the local maxima of the vector-
potential. According to our representation, the
edge-lines connect points where there is strong
directional re-enforcement between edge-tangent
vectors. Symmetry lines can be interpreted as
locations where the magnetic ®eld is perpendicular
to the sampling image plane. Edges are locations
where ®eld lines are tangential to the relevant
sampling plane. When viewed from the perspective
of the dierential structure of the vector-potential,
symmetry lines are locations where the component
of the curl in the image plane vanishes, i.e.
^z ^ r ^
Ax; y; z 0, edges are locations where
the transverse component of the divergence van-
ishes, i.e. r ^z ^
Ax; y; z 0.
Corners, or points of locally maximum boun-
dary curvature, can be viewed as edge-locations
where there is a local symmetry axis associated
with a rapid change in boundary direction. When
viewed from the perspective of our image repre-
sentation, corners therefore correspond to
638 B. Luo et al. / Pattern Recognition Letters 20 (1999) 635±650