Therefore, the fBm index H (and hence D) can be
easily obtained from the slope of this variance
plotted as a function of scale in a log-log plot.
4.2 Local Fractal Dimension
Local fractal dimension is considered as a function
of scales, from (16), we obtain local fractal dimen-
sion as
DQ^) = log, (Var (1,.(n))) ~ og, (var (#, (n) (18)
S. MULTISCALE TEXTURE FEATURE FUSION
Fusion of multiscale texture features is following
feature extraction and is- according to the lateral
inhibition and end-inhibition in neurodynamics. Both
competitive fusion and cooperative fusion are
developed.
5.1 Local Competitive Interactions
Competitive interactions help in noise suppression
and reducing the effects of illumination (Grossberg,
1987; Manjunath, 1993). These steps can be modeled
by non-linear lateral inhibition between features.
Two types of such interactions are identified:
competition between spatial neighbors
with each orientation, and competition between dif-
ferent orientations at each spatial position.
5.1.1 Competition Between Spatial Neighbors with
Each Orientation: A cell of prescribed orientation
excites like-oriented cells corresponding to its
location and inhibits like-oriented cells correspond-
ing to nearby locations at the next processing stage
(Grossberg, 1987). :
Let Y(s,0) be the output of a cell at position s 7 (x, y)
in a given scale with a preferred orientation 0,
1(s,0) be the excitatory input to that cell from the
previous processing stage (texture measures in mult-
iscale analysis), N, be the local spatial neighborhood
of s. These interactions are modeled by non-linear
lateral inhibition between features as
AX (s, 0) « -a,,X(s,0) - I(s,0) LIT (19)
Y(s,0) = g[ X(s, 0)] (20)
where (a, b) are positive weights, g(x) is a non-
linear function such as g(x)= |
(1+exp(-6x))
5.1.2 Competition between different orientations
at Each spatial position: This competition defines a
push-pull opponent process. If a given orientation 0
at position s = (x, y) is excited, then other orienta-
tion Q(0z 9) is inhibited (especially in perpendicu-
lar orientation) and vice versa.
Still, let Y(s,0 be the output of a cell in this step,
the output from previous competition Y(s,0) be the
input /(s,0) to that cell. The competitive dynamics is
represented by
AX (s, 0) 2 —a,, X (5,0) - I(s,0) 2b, XY (s, 9)
Y(s,0)= g[X(s,0)]
(21)
(22)
5.2 Competition Between Scale Interactions
Scale interactions are used for the representation of
end-inhibition property exists among hypercomplex
cells in the visual cortex of mammals. These cells
respond to small lines and edges in their receptive
field, and their response decreases as the length of
lines or edges increases (hence these are often re-
ferred to as end detectors) (Manjunath, 1993). These
cells appear to play an important role in localizing
line-ends and texture boundaries.
If Q,(s,0) denotes the response of such a cell at
position s = (X, y) receiving inputs from two chan-
nels 1 and j (& >a’) with preferred orientation 6,
then
0,(s,0)=g(a1(s,0)- al (s,0) (23)
5.3 Cooperative Fusion
This final stage involves grouping similar orienta-
tions. The cooperative fusion process receives
inputs from the competitive stage and from end-de-
tectors described in local competitive interactions
If. .Z(s,0)
output of this process, then
and scale interactions. represents the
Z(s,0 = g([d(s-s")Y(s', +O,(s', 9))ds"
d(s=(x,y),0) =
(24)
exp[-Qo) Jes cos 0+ ysin 6) + (7x sin 6+ ycos 6) (25)
d(s,0) represents the receptive field of Z(s,0),0 is
the preferred orientation , @ is the corresponding
orthogonal direction, and A is the aspect ratio of
the Gaussian.
1002
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996