7A-3-3
has more practical use than sign function in some cases. In
addition to its simplicity, it provides a smoother function in
comparison to sign function^
c) Sigmoid function
A sigmoid function
f(x,y,k,t)=-lanh(k(\x-)\-t) (8)
satisfies our decision function requirements appropriately. The
shape of the curve near the threshold is approximately that of
the linear function. The curve is generally smooth. Depending
on the parameter k, the first derivative at t could also be
smooth.
Output function
For neuron i, if its charge is Uj that is computed in the energy
minimization, its neuron state output is represented as
v ' = g(u,) =
1
\+e
-HI, IT
(9)
T is the “temperature” (an annealing term) that determines the
speed and quality of the final solution. A very large value of T
will cause neuron values to be 1, while a very small a value
will drive the network to a local minimum state, or a slow
convergence. An annealing process keeps the value of T large
at the beginning and reduces the T value as iteration
progresses. This is important for achieving a global minimum
and a fast convergence.
Initialization
The initial values of neuron states can be chosen randomly as
described in Lin et al. (1991). The network may converge to a
local minimum state. As stated above, an annealing process
may overcome this problem. However, C jk may be calculated
and used as a byproduct to set the initial neuron states as
c l' Yf*’ > w andC) k > w
jeS, jeS {
K=\ C ifjc),>wandc‘ > >()
0, ifCl< 0
(10)
r l N '
where £. is I Cj k > 0] and w=
Combining matched features
After iterations using homomorphism, each neuron reached its
final state V jk ■ Those final states close to 1 yield matches
between corresponding input image features with model
features. However, there is still a need to put the matched
features to form object(s). The following procedure combines
the features under the assumption that there are N features
forming an object.
a) Establish N sets of ={fcll^»l}, i=l,...,/V. Each set
contains all the input features that matched the
corresponding model features.
b) Establish an empty set Q.
c) Set i to be 1.
d) For each k^S i we get m ik =V ik ,keS i if Q is empty,
/ otherwise m v Find the feature k„ that
"Hk ~ ZJ ^¡kjl ^ n
UJ*Q
satisfies \/k p e S t , m iK - m ikp > add (i,kj t0 Q
e) If i=N, one object is recognized and detected, go back to
step a); otherwise, i =/+1 and go back to step b).
3. MULTILAYER HOPFIELD NEURAL NETWORK
The above object recognition is based on separate single layer
neural networks. However, the interrelationship between line
pattern recognition and region pattern recognition adds more
constraints and thus achieves better results. Our objective is to
utilize a two-layer network for truck recognition from aerial
images (Figure 4). The interrelationship is based on the
following realities:
The top of a truck shown in the image is nearly rectangular in
shape, while its shadow has a more complex shape due to the
sun light direction and the truck head shape.
Figure 3. Two-layer (line pattern and region pattern)
Hopfield neural network
The line pattern recognition could yield a match for the top, but
may fail in shadow verification, which is a very important clue;
and the region pattern recognition considers the top and
shadow at the same time, but it does not take full advantages of
line patterns.
Connections among neurons in each single layer are fully
dependent on geometric and photogrammetric constraints and
are fixed before the initial iteration. During iterations the
interconnections between the two layers vary. Let L, denote
layer 1, which is a line pattern layer, and L, denote layer 2,
which is a region pattern layer. We thus have an energy
function