Professional Documents
Culture Documents
Active Contours Without Edges: Tony F. Chan, Member, IEEE, and Luminita A. Vese
Active Contours Without Edges: Tony F. Chan, Member, IEEE, and Luminita A. Vese
2, FEBRUARY 2001
Abstract—In this paper, we propose a new model for active con- the image (the external energy). Observe that, by minimizing
tours to detect objects in a given image, based on techniques of the energy (1), we are trying to locate the curve at the points
curve evolution, Mumford–Shah functional for segmentation and of maxima , acting as an edge-detector, while keeping a
level sets. Our model can detect objects whose boundaries are not
necessarily defined by gradient. We minimize an energy which can smoothness in the curve (object boundary).
be seen as a particular case of the minimal partition problem. In A general edge-detector can be defined by a positive and de-
the level set formulation, the problem becomes a “mean-curvature creasing function , depending on the gradient of the image ,
flow”-like evolving the active contour, which will stop on the de- such that
sired boundary. However, the stopping term does not depend on
the gradient of the image, as in the classical active contour models,
but is instead related to a particular segmentation of the image. We
will give a numerical algorithm using finite differences. Finally, we
For instance
will present various experimental results and in particular some
examples for which the classical snakes methods based on the gra-
dient are not applicable. Also, the initial curve can be anywhere in
the image, and interior contours are automatically detected.
Index Terms—Active contours, curvature, energy minimization,
where , a smoother version of , is the convo-
finite differences, level sets, partial differential equations, segmen- lution of the image with the Gaussian
tation. . The function is positive in
homogeneous regions, and zero at the edges.
In problems of curve evolution, the level set method and in
I. INTRODUCTION
particular the motion by mean curvature of Osher and Sethian
div
(1)
A geometric active contour model based on the mean curva-
Here, , and are positive parameters. The first two terms ture motion is given by the following evolution equation [3]:
control the smoothness of the contour (the internal energy),
while the third term attracts the contour toward the object in div
in (2)
Manuscript received June 17, 1999; revised September 27, 2000. This work
was supported in part by ONR under Contract N00014-96-1-0277 and NSF Con- in
tract DMS-9626755. The associate editor coordinating the review of this man-
uscript and approving it for publication was Prof. Robert J. Schalkoff. where
The authors are with the Mathematics Department, University of Cali- edge-function with ;
fornia, Los Angeles, CA 90095-1555 USA (e-mail: chan@math.ucla.edu;
lvese@math.ucla.edu). is constant;
Publisher Item Identifier S 1057-7149(01)00819-3. initial level set function.
1057–7149/01$10.00 © 2001 IEEE
CHAN AND VESE: ACTIVE CONTOURS WITHOUT EDGES 267
Its zero level curve moves in the normal direction with speed different types of contours, we refer the reader to [8]). In addi-
and therefore stops on the desired tion, our model has a level set formulation, interior contours are
boundary, where vanishes. The constant is a correction term automatically detected, and the initial curve can be anywhere in
chosen so that the quantity div the image.
remains always positive. This constant may be interpreted as a The outline of the paper is as follows. In the next section we
force pushing the curve toward the object, when the curvature introduce our model as an energy minimization and discuss the
becomes null or negative. Also, is a constraint on the area relationship with the Mumford–Shah functional for segmenta-
inside the curve, increasing the propagation speed. tion. Also, we formulate the model in terms of level set functions
Two other active contour models based on level sets were and compute the associated Euler–Lagrange equations. In Sec-
proposed in [13], again using the image gradient to stop the tion III we present an iterative algorithm for solving the problem
curve. The first one is and its discretization. In Section IV we validate our model by
various numerical results on synthetic and real images, showing
the advantages of our model described before, and we end the
paper by a brief concluding section.
Other related works are [29], [10], [26], and [24] on active
in
contours and segmentation, [28] and [11] on shape reconstruc-
where is a constant, and and are the maximum and tion from unorganized points, and finally the recent works [20]
minimum values of the magnitude of the image gradient and [21], where a probability based geodesic active region
. Again, the speed of the evolving curve becomes zero on the model combined with classical gradient based active contour
points with highest gradients, and therefore the curve stops on techniques is proposed.
the desired boundary, defined by strong gradients. The second
model [13] is similar to the geometric model [3], with . II. DESCRIPTION OF THE MODEL
Other related works are [14] and [15]. Let us define the evolving curve in , as the boundary of an
The geodesic model [4] is open subset of (i.e. , and ). In what follows,
inside denotes the region , and outside denotes the
region .
(3)
Our method is the minimization of an energy based-segmen-
tation. Let us first explain the basic idea of the model in a simple
This is a problem of geodesic computation in a Riemannian case. Assume that the image is formed by two regions of ap-
space, according to a metric induced by the image . Solving proximatively piecewise-constant intensities, of distinct values
the minimization problem (3) consists in finding the path of and . Assume further that the object to be detected is repre-
minimal new length in that metric. A minimizer will be ob- sented by the region with the value . Let denote its boundary
tained when vanishes, i.e., when the curve by . Then we have inside the object [or inside ( )],
is on the boundary of the object. The geodesic active contour and outside the object [or outside ( )]. Now let us
model (3) from [4] also has a level set formulation consider the following “fitting” term:
div
(4)
in
in
Because all these classical snakes and active contour models where is any other variable curve, and the constants , ,
rely on the edge-function , depending on the image gradient depending on , are the averages of inside and respec-
, to stop the curve evolution, these models can detect only tively outside . In this simple case, it is obvious that , the
objects with edges defined by gradient. In practice, the dis- boundary of the object, is the minimizer of the fitting term
crete gradients are bounded and then the stopping function
is never zero on the edges, and the curve may pass through
the boundary, especially for the models in [3], [13]–[15]. If the
image is very noisy, then the isotropic smoothing Gaussian This can be seen easily. For instance, if the curve is outside
has to be strong, which will smooth the edges too. In this paper, the object, then and . If the curve is
we propose a different active contour model, without a stopping inside the object, then but . If the curve
edge-function, i.e. a model which is not based on the gradient is both inside and outside the object, then and
of the image for the stopping process. The stopping term is . Finally, the fitting energy is minimized if ,
based on Mumford–Shah segmentation techniques [18]. In this i.e., if the curve is on the boundary of the object. These basic
way, we obtain a model which can detect contours both with or remarks are illustrated in Fig. 1.
without gradient, for instance objects with very smooth bound- In our active contour model we will minimize the above fit-
aries or even with discontinuous boundaries (for a discussion on ting term and we will add some regularizing terms, like the
268 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 10, NO. 2, FEBRUARY 2001
if
if
CHAN AND VESE: ACTIVE CONTOURS WITHOUT EDGES 269
(in the sense of distributions), we express the terms in the energy cases, there are no constrains on the values of and . Then,
in the following way (see also [7]): and are in fact given by
average in
Length average in
and
(7)
in
In this paper, we introduce and use in our experiments the fol-
in lowing regularization of
on (9)
III. NUMERICAL APPROXIMATION OF THE MODEL These distinct approximations and regularizations of the func-
First possible regularization of by functions, as pro- tions and (taking ) are presented in Fig. 3. As
posed in [27], is , both approximations converge to and . A differ-
ence is that has a small support, the interval , while
if
is different of zero everywhere. Because our energy is non-
if convex (allowing therefore many local minima), the solution
may depend on the initial curve. With and , the al-
if
gorithm sometimes computes a local minimizer of the energy,
CHAN AND VESE: ACTIVE CONTOURS WITHOUT EDGES 271
2
Fig. 6. Detection of lines and curves not necessarily closed. Size = 64 64,
(x; y ) = 0 0
(x 32:5) + (y 0 1
32:5) + 30, = 0:02 255 , no
reinitialization, cpu = 2:88 s.
Fig. 9. Object with smooth contour. Top: results using our model without edge-function. Bottom: results using the classical model (2) with edge-function.
compute by the following discretization and linearization where is our solution at time . Then the new
of (9) in will be , such that is obtained at the steady state of (10).
The solution of (10) will have the same zero-level set
as and away from this set, will converge to 1. To
discretize the equation (10), we use the scheme proposed in [22]
and [25].
Finally, the principal steps of the algorithm are:
• Initialize by , .
• Compute and by (6) and (7).
• Solve the PDE in from (9), to obtain .
• Reinitialize locally to the signed distance function to the
curve (this step is optional).
• Check whether the solution is stationary. If not,
and repeat.
We note that the use of a time-dependent PDE for is not
crucial. The stationary problem obtained directly from the mini-
mization problem could also be solved numerically, using a sim-
ilar finite differences scheme.
Fig. 12. Spiral from an art picture. Size = 234 2 191, = 0:000 003 3 1 255 , five iterations of reinitialization, cpu = 108:85 s.
the geometric model (2)], by which the curve cannot detect the In Fig. 10, we validate our model on a very different problem:
smooth boundary. to detect features in spatial point processes in the presence of
CHAN AND VESE: ACTIVE CONTOURS WITHOUT EDGES 275
Fig. 13. Detection of the contours of a plane from a noisy image. Size = 87 2 53, 1t = 0:01, = 0:17 1 255 , (x; y) = 0 (x 0 45) + (y 0 39) +6,
= 2 87
one iteration of reinitialization, cpu : s.
smooth the initial image, even if it is very noisy and in this way,
the locations of boundaries are very well detected and preserved.
By our model, we can detect objects whose boundaries are not
necessarily defined by gradient or with very smooth boundaries,
for which the classical active contour models are not applicable.
Finally, we can automatically detect interior contours starting
with only one initial curve. The position of the initial curve can
be anywhere in the image, and it does not necessarily surround
the objects to be detected. We validated our model by various
numerical results.
Fig. 15. Grouping based on shape identity. In our model, we replaced u from ACKNOWLEDGMENT
Fig. 14 top left, by the curvature of the level curves of u (Fig. 14 top right). Size
= 2 1
64 64, = 0:05 255 , (x; y ) = 0 (x0 32:5) + (y 0 32:5) +
The authors would like to thank J.-M. Morel and G. Sapiro
30:5, 5 iterations of reinitialization, cpu = 10:20 s. for suggesting us the references on Kanisza’s work [8] and Para-
gios–Deriche work [20], [21], respectively, to R. Fedkiw, M. J.
Black, and S. Ruuth for valuable conversations, and especially
to S. Osher for the very useful discussions in the “level set col-
lective” at UCLA. Finally, they wish to thank all three unknown
referees for their general and detailed comments and sugges-
tions, which helped very much to improve the presentation of
the paper.
REFERENCES
[1] G. Aubert and L. Vese, “A variational method in image recovery,” SIAM
J. Numer. Anal., vol. 34, no. 5, pp. 1948–1979, 1997.
[2] S. Byers and A. Raftery, “Nearest-neighbor cluster removal for esti-
mating features in spatial point processes,” J. Amer. Statist. Assoc., vol.
93, no. 442, pp. 577–584, 1998.
Fig. 16. Grouping based on orientation identity. In our model, we replaced u [3] V. Caselles, F. Catté, T. Coll, and F. Dibos, “A geometric model for active
from Fig. 14 bottom left, by the orientation of the normal to the level curves contours in image processing,” Numer. Math., vol. 66, pp. 1–31, 1993.
2 1
of u (Fig. 14 bottom right). Size: 64 64, = 0:025 255 , = 0:02 1 [4] V. Caselles, R. Kimmel, and G. Sapiro, “On geodesic active contours,”
255 , (x; y ) = 0 (x 032:5) + (y 0
32:5) + 30, five iterations of Int. J. Comput. Vis., vol. 22, no. 1, pp. 61–79, 1997.
reinitialization, cpu = 10:25 s. [5] G. Dal Maso, J. M. Morel, and S. Solimini, “A variational method in
image segmentation: existence and approximation results,” Acta Matem-
atica, vol. 168, pp. 89–151, 1992.
Finally, in Fig. 13, the algorithm detects the contours of a [6] A. Dasgupta and A. Raftery, “Detecting features in spatial point pro-
cesses with cluster via model-based clustering,” J. Amer. Statist. Assoc.,
plane from a real noisy image. vol. 93, no. 441, pp. 294–302, 1998.
Of course that our model has its limitations. For instance, it [7] L. C. Evans and R. F. Gariepy, Measure Theory and Fine Properties of
will be interesting to extend the model to the general case of the Functions. Boca Raton, FL: CRC, 1992.
[8] G. Kanizsa, La Grammaire du Voir. Essais sur la perception: Diderot
Mumford–Shah functional. Editeur, Arts et Sciences, 1997.
On the other hand, there are objects which cannot be de- [9] M. Kass, A. Witkin, and D. Terzopoulos, “Snakes: Active contour
tected using the intensity average only. For instance, we show models,” Int. J. Comput. Vis., vol. 1, pp. 321–331, 1988.
[10] S. Kichenassamy, A. Kumar, P. Olver, A. Tannenbaum, and A. Yezzy,
in Fig. 14 two such examples, together with the averages inside “Gradient flows and geometric active contour models,” in Proc. Int.
and outside the objects, which are practically the same (left and Conf. Computer Vision, Cambridge, MA, 1995, pp. 810–815.
middle). One way to overcome this difficulty, would be to use [11] M.-S. Lee and G. Medioni, “Inferred descriptions in terms of curves,
regions and junctions from sparse, noisy binary data,” in Proc. IEEE
other informations from the initial image , like the curvature Int. Symp. Computer Vision, Coral Gables, FL, 1995, pp. 73–78.
(see Fig. 14 top right), or the orientation of level sets (see Fig. 14 [12] C. Lopez and J. M. Morel, “Axiomatization of shape analysis and ap-
bottom right). In this framework, we refer the reader to [12]. plication to texture hyper discrimination,” in Proc. Trento Conf. Surface
Tension Movement Mean Curvature, Berlin, Germany, 1992.
For the results from Figs. 15 and 16, we replaced in our [13] R. Malladi, J. A. Sethian, and B. C. Vemuri, “A topology independent
model by curvature div , and by shape modeling scheme,” in Proc. SPIE Conf. Geometric Methods Com-
orientation respectively (the angle puter Vision II, vol. 2031, San Diego, CA, 1993, pp. 246–258.
[14] , “Evolutionary fronts for topology-independent shape modeling
of the normal to the level curves). Other discriminants may be and recovery,” in Proc. 3rd Eur. Conf. Computer Vision Stockholm,
considered. Sweden, 1994, vol. 800, pp. 3–13.
[15] R. Malladi, J. A. Sethian, and B. C. Vemuri, “Shape modeling with front
propagation: A level set approach,” IEEE Trans. Pattern Anal. Machine
V. CONCLUDING REMARKS AND DISCUSSIONS Intell., vol. 17, pp. 158–175, Feb. 1995.
[16] J. M. Morel and S. Solimini, Segmentation of Images by Variational
In this paper, we proposed an active contour model based Methods: A Constructive Approach. Madrid, Spain: Revista Matem-
on Mumford–Shah segmentation techniques and the level set atica Universidad Complutense de Madrid, 1988, vol. 1, pp. 169–182.
[17] , Segmentation d’Images par Méthode Variationnelle: Une Preuve
method. Our model is not based on an edge-function to stop the Constructive d’Existence. Paris, France: C.R. Acad. Sci. Paris, 1989,
evolving curve on the desired boundary. Also, we do not need to vol. 308, pp. 465–470.
CHAN AND VESE: ACTIVE CONTOURS WITHOUT EDGES 277
[18] D. Mumford and J. Shah, “Optimal approximation by piecewise smooth Tony F. Chan (M’98) received the B.S. degree in
functions and associated variational problems,” Commun. Pure Appl. engineering and the M.S. degree in aerospace engi-
Math, vol. 42, pp. 577–685, 1989. neering, both in 1973, from the California Institute
[19] S. Osher and J. A. Sethian, “Fronts propagating with curvature-depen- of Technology, Pasadena, and the Ph.D. degree
dent speed: Algorithms based on Hamilton–Jacobi Formulation,” J. in computer science from Stanford University,
Comput. Phys., vol. 79, pp. 12–49, 1988. Stanford, CA, in 1978.
[20] N. Paragios and R. Deriche, “Geodesic Active Regions for Texture Seg- He is currently the Department Chair of the De-
mentation,” INRIA RR-3440, 1998. partment of Mathematics, University of California,
[21] , “Geodesic active regions for motion estimation and tracking,” Los Angeles, where he has been a Professor since
INRIA RR-3631, 1999. 1986. His research interests include PDE methods for
[22] E. Rouy and A. Tourin, “A viscosity solutions approach to shape-from- image processing, multigrid, and domain decomposi-
shading,” SIAM J. Numer. Anal., vol. 29, no. 3, pp. 867–884, 1992. tion algorithms, iterative methods, Krylov subspace methods, and parallel algo-
[23] L. Rudin, S. Osher, and E. Fatemi, “Nonlinear total variation based noise rithms.
removal algorithms,” Phys. D, vol. 60, pp. 259–268, 1992.
[24] K. Siddiqi, Y. B. Lauziére, A. Tannenbaum, and S. W. Zucker, “Area and
length minimizing flows for shape segmentation,” IEEE Trans. Image
Processing, vol. 7, pp. 433–443, Mar. 1998. Luminita A. Vese received the M.S. degree in mathe-
[25] M. Sussman, P. Smereka, and S. Osher, “A level set approach for com- matics from the University of Timisoara, Romania, in
puting solutions to incompressible two-phase flow,” J. Comput. Phys., 1993 and the M.S. and Ph.D. degrees in applied math-
vol. 119, pp. 146–159, 1994. ematics, both from the University of Nice-Sophia An-
[26] C. Xu and J. L. Prince, “Snakes, shapes and gradient vector flow,” IEEE tipolis, France, in 1992 and 1996, respectively. Her
Trans. Image Processing, vol. 7, pp. 359–369, Mar. 1998. Ph.D. subject was variational problems and partial
[27] H.-K. Zhao, T. Chan, B. Merriman, and S. Osher, “A variational level differential equations for image analysis and curve
set approach to multiphase motion,” J. Comput. Phys., vol. 127, pp. evolution.
179–195, 1996. She held teaching and research positions at the
[28] H.-K. Zhao, S. Osher, B. Merriman, and M. Kang, “Implicit, nonpara- University of Nice, Nice, France, (1996–1997),
metric shape reconstruction from unorganized points using a variational the University of Paris IX, Paris, France, (1998),
level set method,” UCLA CAM Rep. 98-7, 1998. and the Department of Mathematics, University of California (UCLA), Los
[29] S. C. Zhu, T. S. Lee, and A. L. Yuille, “Region competition: unifying Angeles (1997–2000). Currently she is computational and applied mathematics
snakes, region growing, energy/bayes/MDL for multi-band image seg- Assistant Professor at UCLA. Her research interests include problems of curve
mentation,” in Proc. IEEE 5th Int. Conf. Computer Vision, Cambridge, evolution and segmentation using PDEs, variational methods, and level set
MA, 1995, pp. 416–423. methods.