ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision“, Graz, 2002
zd
e T e e
Vol pon) [rir s (6)
where / is the NxN identity matrix. K is an NxN matrix which
depends on a, f and the shape of the interpolating curves. Ve
gives the location of the element nodes at iteration n.
3. EXTERNAL FORCES CALCULATION
The calculated external force is defined as a function of the
region and the boundary information. We assume that the
image can be modelled as a collection of homogeneous regions
of uniform or slowly varying intensities. Let the image data be
described by a random field Y defined on a rectangular grid, S,
of N points, and the value of Y at a point s € S will be written
as Y,. When necessary the points in S will be explicitly written
as integer pairs (n,m). X will denote the random field of a
given segmentation of Y into uniform regions. Lowercase x and
y will be realizations of the X and Y respectively. The
probability P(Z=z) is written as p(z).
3.1 Statistical Measure
By definition, topographic features could be described by more
than one texture. Indeed, for example, both hardwood and
coniferous trees are included in a vegetation region definition.
Also, these two types of forest can appear separately or in
mixed areas. Thus, the statistical properties of the targeted
area are computed assuming the region of interest is described
by a number of textures being normally distributed. We
propose a mixture estimation process independent but
necessary for the region localization. Indeed, the definition of
the mixture parameters for different type of topographic feature
in an image can be recorded in a database and used as a prior
knowledge. This information is essential to the snake process
but can be defined in an earlier step.
The statistical information is described trough a number k of
Gaussian distributions with the mixture proportion, pj, the
expected value of the argument, u;, and the covariance matrix,
X, with j € [1, k]. The dimension of x; and 2; are related to the
number of bands available in the multi-spectral imagery.
Indeed, the definition of these parameters through n bands
provides 4; vectors of dimension n and 2; matrix of size nxn.
3.1.1 Statistical prior knowledge
The goal is to define the ideal number of textures contained in
a sample region witch describes the data. Recent studies
(Olivier et al., 1999) propose an iterative algorithm for normal
mixture definition. This operation is a two steps iterative
process. It is done using the Expectation — Maximization (EM)
algorithm assuming an adequate initialization of the
parameters values for a given number of distributions. The
objective of the EM algorithm is to maximize the log-
likelihood in an iterative manner. The second step consists of
evaluating the balance between the accuracy of the model and
the number of components. Evaluating the Minimum
Description Length (MDL) of the model provides this measure
as described by Olivier et al., 1999.
The mixture model that minimizes the MDL criterion
corresponds to the appropriate mixture model for the region of
interest.
Since the statistical characteristics of the area are estimated, a
goodness function based on these statistical features can be
derived. Indeed, this function will define the region probability
measure part of the external energy.
3.1.2 Region Probability Measure
Considering a realization of the field Y noted, y, the aim here is
to evaluate p(y|x). The statistical goodness measure is based on
the calculation of a distance, y(y), that evaluates the proximity
of a pixel to the statistical properties of the region. This value
can be defined using the Mahalanobis distance.
Having k textures for a given region, we associate to the
sample y the smallest distance from all k distributions defined
within the region. In order to consider this value in a generic
way, one must take into consideration the number of bands
from the multi-spectral image involved. Since we assume that y
follows a Gaussian distribution, we know that the Mahalanobis
distance follows a x? distribution (Saporta, 1990) with n
degrees of freedom (n being the number of bands for the multi-
spectral image). Thus, we can set a threshold value based on
some confidence value (0.00596 for instance). This threshold
value will serve to normalize the calculated distance for tested
samples and also to identify outliers. Thus, the goodness
function, w(y), can be defined as follow:
min, ne -u) Y. 3 (y —H; )
v(v)=
V,
where y, is a threshold given by the x? table value according
(7)
to the dimension of the data and the desired confidence value.
From Baysian inference, the joint probability distribution,
p(y,0), comprises two parts, a prior distribution p(0) and a
likelihood p(y|0). In this framework, we do not consider the
prior distribution of 0, p(0) in the evaluation of the closest
distribution nor for the goodness function. The idea is to
consider the closest texture according to the Mahalanobis
measure for a candidate as the unique potential.
By observing the distance defined by equation (7), we find that
when it is greater than 1, odds are that this pixel is outside the
region. On the opposite, as the value decreases near zero, the
indication of being inside the region increases. This
information allows the definition of the direction of growth.
Finally, the likelihood p(y|x) when the snake evolves inside the
considered region (i.e., Y(y)<l), is given by a Gaussian
distribution as follows:
px) Sule od (8)
(2x)2V31
For cases, where the snake is outside the considered region
(i.e. w(y)>1), the likelihood is described by a translated form
of the Gaussian given in equation (8). The general case is
defined as follows:
A - 183