angle
ment,
ation
2 list.
y is c,
n the
nged
ed to
ng of
ction
: less
d the
ristic
Hu Xiangyun
and the thresholds of the unit would convergent to a steady state of representation of the pattern. The knowledge of the
recognized object is stored as the weights, I/O characteristic, thresholds and the initial states of the units. Here a
Hopfield neural network is constructed as an optimization way for the extraction of the ‘continuous, smooth and light
ribbon’ from aerial image (see figure 3, § and S, are seed points). In the net, adjacent peak correlation values are
connected each other. The original value of the unit is correlation coefficient, the weight W; is a function of the distance
of the unit (d), difference of the gray at the center of the template matching and average of correlation coefficient of the
two unit (v). According to the road feature, the less d and d, are and the greater v is, the greater W, is, as follow:
W,= C,D/d+C v+C d, (1)
Here C,, C, and C, are weight coefficient of the items, we set their value 0.45, 0.25 and 0.3. D is matching interval. The
I/O characteristic of the unit is defined by a sigmond function:
fi luo exp( -(3- Wi — 1))) (2)
I is the threshold of the unit, it is direct ratio of the distance between the unit and the fitted curve. The experimental
result shows that the convergence of the net (the states of all units don't change) is right output even there is heavy
noise. For example, there are not active output units where the road is shadowed by trees. The performance of Hopfield
neural network assures the convergence and our I/O function of the unit guarantees fast convergence.
4 EXPERIMENTAL RESULTS
We have integrated the algorithm into a digital photogrammetric workstation — VirtuoZo™. Using the system, many
(a) (b) (c)
Figure 4. Experimental result of one road extraction
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 997