ACTIVE CONTOURS FOR VECTOR-VALUED IMAGES 131
gradientofu
0
)attractsthe curvetoward the boundary.Forinstance,thismodelcannothandle
automatic topology changes of the contour, and also it depends on the parameterization of
the curve.
In problems of curve evolution, including snakes and active contours, the levelset method
of Osher and Sethian [3] has been used extensively because it allows for automatic topology
changes, cusps, and corners; moreover, the computations are made on a fixed rectangular
grid. Using this approach, geometric active contour models, using a stopping edge-function,
have been proposed in [4] and also in [5–7]. These models are based on the theory of
curve evolution and geometric flows and in particular on the mean curvature motion of
Osher and Sethian [3]. The evolving curve moves by mean curvature, but with an ex-
tra factor in the speed, by the stopping edge-function. Therefore, the curve stops on the
edges, where the edge-function vanishes. A typical example of edge-function used is given
by
g(|∇u
0
|) =
1
1 +|∇(G
σ
∗u
0
)|
2
,
where g is a positive and decreasing function such that lim
t→∞
g(t) =0. The image u
0
is
first convolved with the Gaussian G
σ
(x, y) =σ
−1/2
exp
−|x
2
+y|
2
|/4σ
, especially for the cases
where u
0
is noisy. But in practice, g is never zero on the edges, and therefore the evolving
curve may not stop on the desired boundary. To overcome this problem, a new model has
been proposed in [8] as a problem of geodesic computation in a Riemannian space, accord-
ing to the metric g. The associated Euler–Lagrange equation, in level set formulation, has
a new additional gradient term (compared with the previous geometric models [4–7]). This
term increases the attraction of the evolving curve toward the boundary of the object and
is of special help when the boundary has high variations on its gradient values. For another
related approach, see also [9].
These models use the gradient of a smoother version of the image u
0
to detect edges.
But, if the image is noisy, the smoothing in the edge-function has to be strong, thus blurring
edge features, or a preprocessing has to be implemented, to remove the noise.
In contrast, the Chan–Vese (C-V) active contour model without edges proposed in [1]
does not use the stopping edge-function g to find the boundary. Instead, the stopping term
is based on Mumford–Shah segmentation techniques. This model has several advantages:
it detects edges both with and without gradient (see [10] for a discussion on edges without
gradient, called cognitive edges); it automatically detects interior contours; the initial curve
does not necessarily have to start around the objects to be detected and instead can be placed
anywhere in the image; it gives in addition a partition of the image into two regions, the
first formed by the set of the detected objects, while the second one gives the background;
finally, there is no need for an a priori noise removal.
Thepreviously describedmethodshavebeendevelopedtodetectobjectsinasingle image,
but what happens when we have several different registered images of the same object? This
occurs in multispectral images taken at different wavelengths, in medical images taken by
different equipment (i.e., PET, MRI, and CT), in color images, or in textured images. Each
image channel may have signal characteristics that can be combined with other channels to
enhance contour detection.
Severalmodelsofrestoration,edge detection,andalsoactivecontourshavebeenproposed
for vector-valued images. Forrestoration of color images, we mention the works in [11–13].