Regularized Laplacian Zero Crossings as Optimal Edge Integrators
R. KIMMEL A.M. BRUCKSTEIN
Department of Computer Science
Technion–Israel Institute of Technology
Technion City, Haifa 32000, Israel
Abstract
We view the fundamental edge integration problem for object segmentation in a geometric vari-ational framework. First we show that the classical zero-crossings of the image Laplacian edgedetector as suggested by Marr and Hildreth, inherently provides optimal edge-integration with re-gard to a very natural geometric functional. This functional accumulates the inner product betweenthe normal to the edge and the gray level image-gradient along the edge. We use this observation toderive new and highly accurate active contours based on this functional and regularized by previouslyproposed geodesic active contour geometric variational models.
1. Introduction
Edge integration for segmentation is an old, yet still very active area of research in low-level image
analysis. Textbooks in computer vision treat edge detection and edge integration as separate topics, the
first being considered one of labelling edges in the image to be followed by a process of integrating
the local “edges” into meaningful curves. In fact one may view basic edge detection as a process of
estimating the gradient of the image, i.e. computing at each pixel���������
the values ��������and � � ������
by using the values of � � ������over a neighborhood � � ������
of� ������
and designating as edges the places
where the length of the gradient vector estimate �� � ����� ��� � � exceeds some threshold value.
The more advanced edge detectors such as those proposed by Marr and Hildreth [7] attempt to locate
points or curves defined by local maxima of the image gradient. The Marr Hildreth proposal for edge
1
Tec
hnio
n - C
ompu
ter S
cien
ce D
epar
tmen
t - T
echn
ical
Rep
ort
CIS
-200
1-04
- 20
01
detection yields curves that delineate the zero crossing of the Laplacian operator applied to a smoothed
version of the image input. The smoothing proposed is via a Gaussian convolution operator and its width
is a parameter that can be varied providing the opportunity to do scale space processing and “vertical”
integration on the zero-crossing curves.
In this paper we propose to regard the edge detection and integration process as a way to determine
curves in the image plane that pass through points where the gradient is high and whose direction best
corresponds to the local edge direction predicted by the estimated gradient.
Indeed, if we somehow estimate the gradient field �� ���������� � � ������ � based on considering � ���������for
each pixel� ������
over some neighborhood ��� � ������, where � is a size parameter, we shall have at each
point a value, given by the intensity of the gradient� � ����������� ��� � ������ �"!# , that tells us how likely
an edge is at this point and, if an edge exists, its likely direction will be perpendicular to the vector�� � �������� � � ������ � . It is therefore natural to look for curves in the image plane, $ ��%"� �&� �� %"��������%'� � , that
pass through points with high intensity gradients with tangents agreeing as much as possible to the edge
directions there. Thus we are led to consider the following functional, evaluating the quality of $ ��%'�as
an edge-curve candidate,(*) � $ � %"�+� � ,.-/10324 �� � $ ��%'�+��� � � $ � %"� � �"5768:9 ;< ; 9=>@?BA $ � %"�A % CED�FG A %
where 0 � 5 � is some monotonically increasing function. Here, the inner product of� �H�I�� � �J� with the
normal to $ � %"�, given by KL � %"� � 68 9 ;< ; 9
=> ?MA $ ��%'�A % C�D �where $ ��%"�
is an arclength parameterized curve, is a measure of how well $ � %"�is locally tracking an
edge. Indeed we want $ ��%'�to pass at high gradient locations in the edge direction, and hence the inner
product of its normal with the estimated gradient of � should be high, indicating both alignment and
considerable change in image intensity there. This inner product will also be proportional to the gradient
magnitude, since �� � $ ��%"� ��� � � $ � %"�+� � 68 9 ;< ; 9=> ? A $A % C D �ON � �PNQ5SRUT'V � W����
2
Tec
hnio
n - C
ompu
ter S
cien
ce D
epar
tmen
t - T
echn
ical
Rep
ort
CIS
-200
1-04
- 20
01
whereW
is the angle between the outward pointing normal
KL to $ � %"�and the gradient direction.
The functionals(X)
measure how well an arclength parameterized curve of length Y approximates an
edge in the image plane. Our task, of course is to determine several most probable edge curves in the
image plane. We shall do so by determining curves that locally maximize these functionals.
Suppose first that we are considering closed contours $ � %"�, and that 0 ��ZX� � Z
. Then, we have that( � $ ��%'�+� �\[^]E_a`cbedf � $ ��%"� � AA % ����%'� < � � $ ��%'�+� AA % ���%'��g A %��and Green’s theorem yields( � $ ��%'�+� � ,e,.h�i dkjj � � � ������E� jj � � ������ g A � A �E�where l ]
is the region inside $ � %"�. But, recalling that � � ������
is an estimate of mmon � � ������and ��������
an
estimate of mmop � � ������we have( � $ ��%'�+�rq� ,e, h�i d j �j � � � � ������f� j �j � � � � ������ g A � A �q� ,e,.h�i ��s � � ������ � A � A ��t
Therefore, the functional that we want to maximize is the integrated Laplacian over the area enclosed by$ � %"�. This means that if we have an area where the Laplacian is positive
( � $ � %"� �should expand from
within this area to the places wheres � � ������
becomes zero and subsequently changes sign. This shows
that optimal edge-curves in the sense of maximizing( � $ � %"� �
are the zero crossings of the Laplacian.
If we initialize $ � %"�as a small circular “bubble” at a place where
s � � ������is positive and then let$ � %"�
evolve according to a rule that implements a gradient descent in conjunction with the functional( � $ � %"� ���i.e. we implement AA"u $ � %�� u � �Iv ( � $ � %�w u � �v $ �
the curve $ � %�� u �will expand in time
uto the nearest zero-crossing curve of the input image Laplacian.
Therefore, we have obtained a beautiful interpretation of the classical Marr-Hildreth edge detection
method [7]. The zero-crossings of the Laplacian are curves that best integrate the edges, in the sense
of our functional(X) � $ � %"� �
with 0 ��ZX� � Z, if we wish to do so based on gradients estimated for the
(smoothed) input image � � ������. While this fact is pedagogically very pleasing, it does not alleviate the
3
Tec
hnio
n - C
ompu
ter S
cien
ce D
epar
tmen
t - T
echn
ical
Rep
ort
CIS
-200
1-04
- 20
01
notorious over sensitivity properties of this edge-detector which in noisy images yields lots of false edge
curves. However, we shall show here that this insight provides the basis for a new and practical active
contour process which enhances and improves upon the previously designed such methods for image
segmentation.
We next present the full derivation of the variational results leading to the new edge integration pro-
cesses and then show the performance of the resulting algorithms.
2. Closed Active Contours: Derivation
Motivated by the classical ‘snakes’ [4], geometric active contours [5, 1], and finally the ‘geodesic active
contours’ that were shown in [2] to be related to the ‘snakes’, we search for simple parametric curves
in the plane that map their arclength interval � 9 � Yx� to the plane, such that $zy{� 9 � Yx�7| IR � , or in an
explicit parametric form $ ��%"� �}� ���%'�����E� %"� � . Here%
is the arclength parameter, and we have the relation
between the arclength%
and a general arbitrary parameterization ~ , given byA % � d A �� ~ �A ~ g � � d A �E� ~ �A ~ g � A ~��ON�$*�^N A ~ tWe define, as usual,
KL ����and
Kuto be the unit normal, the curvature, and the tangent of the curve $ .
We have that� KL �\$ `c`
, and
Ku ��$ ` �\$ �S� N�$ � N . As described in the introduction, consider the geometric
functional (�) � $ � �\[ -/10 �+�K� � KL� � A %�t
This is an integration along the curve $ of a function 0 defined in terms of a vector field
K� ���� � �������� � � ������ � ,where for example we can take
K� � � � �������� ����� p � � n � as the gray level image gradient. Our goal is to
find curves $ that minimize the above geometric functional.
In a general parametric form, we have the following re-parameterization invariant measure(*) � $ � �\[��/�0 �+�K� � KL� � N�$*�^N A ~ t
Define,Z��r� K� � KL� . The Euler Lagrange (EL) equations v ( ) � $ � � v $�� 9 should hold along the ex-
4
Tec
hnio
n - C
ompu
ter S
cien
ce D
epar
tmen
t - T
echn
ical
Rep
ort
CIS
-200
1-04
- 20
01
tremum curves, and for a closed curve these equations arev (�) � $ �v $ � 24 mmop <��� � mm�p �mmon <z�� � mm�n � FG 0 � Z*� N�$*�^N �or in a more compact form v (�) � $ �v $ ��d@jj $ < AA ~ jj $ � g 0 � Z*� N�$ � N �where we use the shorthand notation j � j $\��� j � j ��� j � j � � D and j � j $*����� j � j � � � j � j � ��� D . In case of
an open curve, one must also consider the end points and add additional constraints to determine their
optimal locations.
Before we work out the general 0 � Z*�case let us return to the simple example discussed in the intro-
duction, where 0 � Z*� � Z. In this case we have that( � $ � � [��/ � K� � KL� N�$*�^N A ~� [��/O�
K� � � < � � ��� � �N�$*�^N � N�$*�'N A ~� [��/ � < � �B ��� �M� � A ~ tThe EL equation for the
�part is given byv (v � � d jj � < AA ~ jj � � g � < � � �.� � � �� < � � p ��� � � p < AA ~ �� < � �B p ��� �B� p < � p � � < � n � �� < � � � p � � n � � < � �B�^��� � K� ��t
In a similar way, the�
part of the EL equations is given byv (v � � dkjj � < AA ~ jj � � g.� < � � �.� � � �� < � �B n �.� �B� n � AA ~ � < � �B n �.� �B� n � p � � � n � �� � � � p � � n � � � �B����� � K� ��t5
Tec
hnio
n - C
ompu
ter S
cien
ce D
epar
tmen
t - T
echn
ical
Rep
ort
CIS
-200
1-04
- 20
01
Since the EL equations are derived with respect to a geometric measure, we can use the freedom of
reparameterization for the curve $ , divide by N�$��^N , and obtain the “geometric EL equation:” v ( � v $@��^��� � K� � KL , and for
K� � � � � ������we have v ( � v $\� s � KL , where
s � � � p�p � � n�n is the usual Laplacian
operator. It is obvious from this that the geometric EL condition is satisfied along the zero crossing
curves of the image Laplacian, which as described above explains the Marr-Hildreth [7, 6] edge detec-
tor from a global-variational point of view. Below, we shall extract further insights and segmentation
schemes from this observation. We note that heuristic non-variational flows on vector fields were pre-
sented in [13, 10]. In a recent related result, introduced by Vasilevskiy and Siddiqi [12], alignment with
a vector field is used as a minimization criteria for segmentation of complicated closed thin structures in
3D medical images.
As a second example we consider 0 � Z*� ��N Z NJ�@� Z � . The EL is given byv (v $ � � K� � KL�N � K� � KL� N �^��� �K� � KL� V��� "¡ �+� K� � KL�� � ����� � K� � KL �
and for
K� � � � we have v ( � v $¢��V��� "¡ � � � � � KLf� ��s � KL . The new term V£�� "¡ � ��� � � KL¤� � , allows the model
to automatically handle changing contrasts between the objects and the background. For example, it
handles equally well an image of dark objects on bright background and the negative of this image.
Now, we are ready to pursue the general case for 0 ��ZX�in the functional
(�) � $ �(where
Z � � K� � KL� ).We shall use often the following readily verified relationships,AA ~ � N�$ � N AA %j N�$ � Nj $*� � KuA KLA % � < � KuA KuA % � � KLA 0 � Z*�A % � 0�¥ Z ` � � � 0 �
Ku �A ZA % � � K� ` � KL� �¦� K� � KL ` � � � K� ` � KL� < ��� K� � Ku �j Zj $*� � < N�$ � N�§ � �K� � Ku � KL �
6
Tec
hnio
n - C
ompu
ter S
cien
ce D
epar
tmen
t - T
echn
ical
Rep
ort
CIS
-200
1-04
- 20
01
and that d jj $ < AA ~ jj $*� g.� Z N�$*�^N � � N�$*�'N¨����� � K� � KL tUsing these relations we havev ( ) � $ �v $ � d jj $ < AA ~ jj $*� g 0 � Z*� N�$*�^N� 0 ¥ Z ] N�$ � N < AA ~3© 0 ¥ Z ] � N�$ � N � 0
Ku�ª� 0 ¥ Z ] N�$*�'N < AA ~ © 0 ¥ Z ] � N�$*�^N � 0 ¥ Z
Ku < 0 ¥ ZKu � 0
Ku ª� 0�¥ Z ] N�$*�'N < AA ~ © 0�¥ Z ] � N�$*�^N � 0�¥ Z
Ku ª < AA ~ © < 0�¥ ZKu � 0
Ku ª� 0 ¥ Z ] N�$*�'N < 0 ¥ AA ~3© Z ] � N�$*�^N �.Z Ku ª< 0 ¥¨¥ Z � © Z ] � N�$ � N ��Z Ku ª < AA ~ © < 0 ¥ Z
Ku � 0Ku ª� 0�¥ N�$ � N¨����� � K� � KL < 0�¥¨¥ N�$ � N Z `¬« < � K� � Ku � KL ��Z Ku�< N�$ � N AA % © < 0 ¥ Z
Ku � 0Ku ª� N�$*�^N « < 0 �®� 0�¥ � ����� � K� ���� K� � KL�� �f�� 0�¥¨¥ � �
K� ` � KL�� � K� � Ku �< ��� K� � Ku � � � KL �some additional tangential components
tHere we used the shorthand notations
Z ] � ��Z �@� Z p ��Z n � , andZ ] � � � ] � Z .
2.1. Gradient Descent via Level Set Formulation
In order to determine optimal curves in the plane, we need to solve numerically the EL equations. Here
we shall follow the “geodesic active contour philosophy,” see [2], and design a curve evolution rule that
is given by $e¯� v (�) � $ �v $ tThis is a gradient descent rule with respect to the chosen cost functional, and in this flow one can consider
only the normal components of v (e) � $ � � v $ , since tangential components have no effect on the geometry
the propagating curve. Next, we can embed the curve in a higher dimensional ° ���������function, which
7
Tec
hnio
n - C
ompu
ter S
cien
ce D
epar
tmen
t - T
echn
ical
Rep
ort
CIS
-200
1-04
- 20
01
implicitly represents the curve $ as a zero set, i.e., $±�&²³� ���� �´yx° �������� � 9Jµ . In this way, the well
known Osher-Sethian [8, 11] level-set method can be employed to implement the propagation.
Given the curve evolution equation $ ¯ ��¶ KL , its implicit level set evolution equation reads° ¯ ��¶eN � °eN tThe equivalence of these two evolutions can be easily verified using the chain rule and the relationKL � � ° � N � °eN , ° ¯ � ��� ° � $ ¯ � � � � ° � ¶ KL�� ��¶ � � ° � � °N � °¬N � �¦¶¬N � °¬N tWe readily have that Ku � ·� °N � °¬N � � < ° n � ° p �N � °eN� � �^����d � °N � °eN gK� ` � �� ` � � ` �f��� � � � Ku � �Q��� � � Ku � �V��� "¡ � � K� � KL�� � � V��� ¸¡ � � K� ��� ° � ��tThereby, the explicit curve evolution as a gradient descent flow for 0 � Z*� ��N Z N is given by$ ¯ �¹V£�� "¡ �+� K� � KL� ��s � KL �for which the implicit level set evolution is given by° ¯ �¹V£�� "¡ �+� K� ��� ° � ��s �fN � °¬N t3. Open Active Contours for Optimal Edge Integration
3.1. The Fua-Leclerc Geometric Model
Fua and Leclerc in [3], were first to propose a geometric model for motion of open curves in the image
to optimize an “edge” finding functional. We shall first describe the Fua-Leclerc functional and then
replace the “geodesic active contour” part of it with our new edge integration quality measure. LetY � $ � � , �/ N�$*��N A ~ �8
Tec
hnio
n - C
ompu
ter S
cien
ce D
epar
tmen
t - T
echn
ical
Rep
ort
CIS
-200
1-04
- 20
01
be the arclength of an open curve $ � ~ � . Adding the variation º � ~ � to the curve, such that »$ � ~ � �$ � ~ ��½¼ º � ~ � , differentiating w.r.t.¼, and letting
¼go to zero, yieldsY ¥ � $ � � < ,.-/ � KL º A %�� º � Y � Ku � Y � < º � 9 �
Ku � 9 ���where
%is the arclength parameter. Also, following Fua and Leclerc, considerY¾ � $ � � ,.-/1¿ � $ � %"� � A %J�
where ¿ is some suitably defined “edge indicator” function, for example ¿ � ������ � ; � � N � �fNÀ� � ; � . The
first variation of Yx¾ � $ �can be easily shown to be given byY ¥¾ � $ � � ,.-/ � � � ¿ �
KL�� < � ¿ �KL º A %�� º � Y � ¿ � $ � Y � � Ku � Y � < º � 9 � ¿ � $ � 9 � �
Ku � 9 ��tThe Fua-Leclerc functional is defined as ( ¾ � $ � � Y¾ � $ �Y � $ � t
Computing the first variation, we have thatv ( ¾v $ � Y¬Y ¥¾ < Y ¥ Y¾Y � � 9 �should hold for any º . Therefore, the following conditions must be satisfied,YeY ¥¾ � Y ¥ Y¾ �or explicitly, Á�Â�Ã^Ä -/¹Å ÆcÇ�ÈEɸÊË�ÌfÍÏÎ ÈJÐ"ÊË^Ñ�Ò"ÓxÔÕÑ Å Á ÐÖÈ�ÅØ×ÙÅ Á Ð�Ð ÊÚ Å Á Ð Í�Ñ ÅÜÛSÐÖÈPÅ�×ÙÅÝÛSÐ�Ð ÊÚ ÅÝÛSÐ Þß Á ¾  Ã^Ä -/ ÍeÎ ÊËEÑ�Ò"ÓxÔÕÑ Å Á Ð ÊÚ Å Á Ð ÍàÑ ÅÜÛSÐ ÊÚ ÅÜÛSÐ ÞÏáThus, we should verify the following necessary conditions for a local extremum to hold for any º ,,.-/ � Y � ��� ¿ �
KL�� < � ¿ �x� Y¾ �f� KL º A % � 9Y�º � Y � ¿ � $ � Y �+� Ku � Y � � Y¾âº � Y � Ku � Y �Y�º � 9 � ¿ � $ � 9 � �Ku � 9 � � Y¾âº � 9 �
Ku � 9 ��t9
Tec
hnio
n - C
ompu
ter S
cien
ce D
epar
tmen
t - T
echn
ical
Rep
ort
CIS
-200
1-04
- 20
01
Therefore, the geometric conditions that must be met along the curve and at its end points, are:d7d Y�¾Y < ¿ gã�ä����� ¿ �KL� g KL � 9¿ � $ � 9 � � � Y¾Y¿ � $ � Y �+� � Y¾Y t
We can use these conditions to guide a gradient descent process for an active contour evolution toward
the local minimum of the Fua-Leclerc functional. To do that we apply the following evolution equation
along the curve and at its end points,$ ¯ ��d � ¿ < � � ¿ �KL�� < Y¾Y ��g KL t
The first two terms depict the geodesic active contour ([2]) model, while the third term directs the curve
to gain length by applying the inverse geometric heat equation at points where ¿ � $ ��%'�+�æå Ye¾ � Y . We
still need to design the motion of the end points. Consider the end point $ � Y �. The curve should reduce
its length if ¿ � $ � Y � ��ç Y�¾ � Y , in which case the end point should move along the tangent < Ku � Y �. Hence,
for example, we can use the following evolution rules at the end points:$e¯ � 9 � � � Y ¿ � $ � 9 �+� < Y ¾ � Ku � 9 �$ ¯ � Y � � � Y¾ < Y ¿ � $ � Y �+� � Ku � Y ��t3.2. Our Optimal Edge Integration
We propose to use our measure, Y ) �Oè 0 � Z*� A %instead of Y�¾ , in the Fua-Leclerc functional. Here we
compute the evolution equations that propagate the open curve $ towards a maximum of the functional( ) � $ � � Y )Y tTherefore, we are searching for éSê ]ìë éîí (*) � $ �
. The quantity Y in this maximization process, penalizes
the length of the curve, i.e. it plays a role opposite from its role in the minimization of the Fua-Leclerc
functional.
10
Tec
hnio
n - C
ompu
ter S
cien
ce D
epar
tmen
t - T
echn
ical
Rep
ort
CIS
-200
1-04
- 20
01
We now use the v (X) � v $ expression developed in the previous sections for the general 0 � �K� � KL� �
closed curve case. We have thatY ¥ ) � $ � � ,.-/ º v (*)v $ A %ï� º « 0Ku < 0 ¥ �
K� � Ku � KL {ððð -/ tUsing these conditions in the Fua-Leclerc formulae yield along the curve,Y v (*)v $ � Y ) � KL � 9 �and Y � 0
Ku < 0 ¥ �K� � Ku � KL � � Y ) Ku �
at the end points $ � Y �and $ � 9 � . For 0 ��ZX� �ON Z N , the gradient descent flow along the curve is given by$e¯±� V��� ¸¡ � ��� � � KL¤� ��s � KL � Y )Y � KL$ ¯ � 9 � � � Y 0 < Y ) � Ku < 0�¥ �
K� � Ku � KL$ ¯ � Y � � � Y ) < Y 0 �Ku � 0 ¥ �
K� � Ku � KL t3.3. A Simpler Formulation for the New Optimal Edge Integration
Functionals that involve ratio of two integral measures, like the Fua-Leclerc functional, require integra-
tion along the contours for a proper gradient descent flow. Integral parts are present in the EL equations
which require computationally intensive global integration procedures for the computation of the proper
flow. Recall however that our goal is to maximize(�) � $ �
on one hand, that leads to long curves, while
also penalizing the length of the curve on the other hand. We shall therefore consider the following
alternative functional that would also realize these goals,»( ) � $ � �¹Y ) < Y tThe EL equations in this case are given by
9 �Iv (*) � $ �v $ ��� KL �11
Tec
hnio
n - C
ompu
ter S
cien
ce D
epar
tmen
t - T
echn
ical
Rep
ort
CIS
-200
1-04
- 20
01
along the curve, and
9 � � 0 < ; �Ku < 0�¥ �
K� � Ku � KL �at the end points. The motivation for the tangential term at the end point is obvious, it either extends of
shrinks the curve. The normal term pulls it from running parallel to the vector field and directs the end
point towards the center of the edge (where 0 ¥ should be zero). These two components define the motion
at the end points. For 0 � Z*� ��N Z N and
K� � � � we have$ ¯ � V£�� "¡ �+� � � � KLf� �ñs � KL �.� KL$ ¯ � 9 � � � N � � � � KLf� N < ; �Ku < V£�� ¸¡ � ��� � � KLf� ����� � � Ku � KL$ ¯ � Y � � � ; < N � � � � KLf� N � Ku � V£�� "¡ �+� � � � KL� ����� � � Ku � KL t
4. Simulation Results
We tested the edge integration methods discussed in this paper on two simple examples. The first pre-
sented segmentation examples shown are not typical for active contours and could be easily processed
with less sophisticated methods. However, they capture the difficulties of the existing active contour
models and therefore are useful for comparison of the different methods.
In all, the ”closed contour” cases we started from the image frame as the initial contour, and applied
a multi-resolution coarse to fine procedure, as in [9], to speed up the segmentation process.
Figure 1 shows the advantage of the Laplacian model in cases where only the gradient is affected,
the Laplacian being invariant to an additive intensity plane, as well as in cases where the Laplacian is
also changed by a constant when a parabola was added to the intensity surface. Figures 2 and 3 clearly
exhibit the segmentation advantages of the Laplacian active curve model as a core with the geodesic
active term as a regularization. We here used the functional( � $ � �¹[�] 0 ��ZX� A % <kò [³] ¿ � � $ � %"� � A % <ôó ,�h i ¿ � �������� A � A ���where ò and ó are small positive constants, 0 ��ZX� �ON Z N , and
Z � � � � � KLf� . ¿Jõ , ö*� ; ��÷ are edge indicator
functions with lower values along the edges. In this case, the gradient descent flow for maximizing( � $ �
12
Tec
hnio
n - C
ompu
ter S
cien
ce D
epar
tmen
t - T
echn
ical
Rep
ort
CIS
-200
1-04
- 20
01
Figure 1: Top: Synthetic images with a tilted intensity plane added to the original image on the left, and
a parabola intensity surface added to the original image on the right. Middle row: Segmentation results
with the GAC model. Bottom: Segmentation results with the Laplacian model.
13
Tec
hnio
n - C
ompu
ter S
cien
ce D
epar
tmen
t - T
echn
ical
Rep
ort
CIS
-200
1-04
- 20
01
Figure 2: Left: map input image. Right: The active contours results, in which the upper row shows
the geodesic active contour results, middle row presents the results of the Laplacian active contour, and
bottom row is a combination of the two.
is given by $ ¯ � � V£�� ¸¡ � ��� � � KL�� ��s � � ò � ¿ � � < ��� ¿ � �KL�� � <½ó ¿ � �
KL tThe level set formulation for this flow is°E¯ø��dEV£�� "¡ � ��� � ��� ° � ��s � � ò �����ùd ¿ � � °N � °¬N g <½ó ¿ � g N � °¬N t
Next, we applied our open contour model for edge integration on similar images, but here we started
with short contour segments that expanded and locked onto boundaries, if such existed in the vicinity of
the initialized contours. If no boundaries are detected locally, the contour segments shrink and eventually
disappear. The numerical implementation for the open contour case is an explicit ”marker-points” based
model which was easier to program in this case. At each iteration the marker-points are re-distributed
along the contour to form equi-distant numerical representations of the contour. A simple monitoring
procedure, removes a marker point when successive marker points get too close to one another, and adds
a new marker point in the middle of two successive marker points when the distance between them gets
14
Tec
hnio
n - C
ompu
ter S
cien
ce D
epar
tmen
t - T
echn
ical
Rep
ort
CIS
-200
1-04
- 20
01
larger than a given threshold. The examples show how initial segments expand and deform until they
lock onto the boundaries of rather complex shapes. See Figures 4 and 5.
5. Conclusions
In this paper we proposed to incorporate the directional information that is generally ignored when de-
signing edge integration methods in a variational framework. Simulations that were performed with
the newly defined edge integration processes amply demonstrated their excellent performance as com-
pared to the best existing edge integration methods. Our extended active contour models are just a few
examples of the many possible combinations of geometric measures. Other functionals that could be
considered to either open or closed curves are Y ) < Y ¾ , or è 0 ��ZX�ñú�� N � �PN � A % , for an edge indicator func-
tion likeú � ; < ¿ � N � �fN � or
ú^� N � �fN � �üû N � �PN � � ; . For closed curves, the(e)
part is most effective
when the curve is close to its final location, therefore, the functional ý{� 0 � Z*��� ; < ¿ � < ¿ � A % could also be
considered.
Acknowledgments
We thank Evgeni Krimer and Roman Barsky for implementing and testing the open active contours
models.
References
[1] V. Caselles, F. Catte, T. Coll, and F. Dibos. A geometric model for active contours. Numerische
Mathematik, 66:1–31, 1993.
[2] V. Caselles, R. Kimmel, and G. Sapiro. Geodesic active contours. IJCV, 22(1):61–79, 1997.
[3] P. Fua and Y. G. Leclerc. Model driven edge detection. Machine Vision and Applications, 3:45–56,
1990.
15
Tec
hnio
n - C
ompu
ter S
cien
ce D
epar
tmen
t - T
echn
ical
Rep
ort
CIS
-200
1-04
- 20
01
[4] M. Kass, A. Witkin, and D. Terzopoulos. Snakes: Active contour models. International Journal of
Computer Vision, 1:321–331, 1988.
[5] R. Malladi, J. Sethian, and B. C. Vemuri. Shape modeling with front propagation: A level set
approach. IEEE Trans. on PAMI, 17:158–175, 1995.
[6] D. Marr. Vision. Freeman, San Francisco, 1982.
[7] D. Marr and E. Hildreth. Theory of edge detection. Proc. of the Royal Society London B, 207:187–
217, 1980.
[8] S. J. Osher and J. Sethian. Fronts propagating with curvature dependent speed: Algorithms based
on Hamilton-Jacobi formulations. J. of Comp. Phys., 79:12–49, 1988.
[9] N. Paragios and R. Deriche. Geodesic active contours and level sets for the detection and tracking
of moving objects. IEEE Trans. on PAMI, 22(3):266–280, 2000.
[10] N. K. Paragios, O. Mellina-Gotardo, and V. Ramesh. Gradient vector flow fast geodesic active
contours. In Proceedings ICCV’95, Vancouver, Canada, July 2001.
[11] J. Sethian. Level Set Methods: Evolving Interfaces in Geometry, Fluid Mechanics, Computer Vision
and Materials Sciences. Cambridge Univ. Press, 1996.
[12] A. Vasilevskiy and K. Siddiqi. Flux maximizing geometric flows. In Proceedings ICCV’95, Van-
couver, Canada, July 2001.
[13] C. Xu and J. Prince. Snakes, shapes, and gradient vector flow. IEEE Trans. IP, 7(3):359–369,
1998.
16
Tec
hnio
n - C
ompu
ter S
cien
ce D
epar
tmen
t - T
echn
ical
Rep
ort
CIS
-200
1-04
- 20
01
Figure 3: Top: Geodesic active contour results. Middle: Laplacian active contours. Bottom: Using the
geodesic active contours as a regularization for the Laplacian active contour.
17
Tec
hnio
n - C
ompu
ter S
cien
ce D
epar
tmen
t - T
echn
ical
Rep
ort
CIS
-200
1-04
- 20
01
time=0 time=50 time=100 time=150
time=200 time=250 time=300 time=350
time=400 time=450 time=500 time=550
Figure 4: Open geometric Laplacian active contours: The initial small diagonal lines (top left frame)
deform and either shrink and vanish or extend along the boundaries of the objects and capture their
shapes.
18
Tec
hnio
n - C
ompu
ter S
cien
ce D
epar
tmen
t - T
echn
ical
Rep
ort
CIS
-200
1-04
- 20
01
time=0 time=50 time=100 time=150
time=200 time=250 time=300 time=350
time=400 time=450 time=500 time=550
Figure 5: Open geometric Laplacian active contours: The small curves (top left frame) extend along the
boundaries and capture most of the outer contours of the symbols.
19
Tec
hnio
n - C
ompu
ter S
cien
ce D
epar
tmen
t - T
echn
ical
Rep
ort
CIS
-200
1-04
- 20
01
Top Related