OM model. The
ron grid and an
about the object
the image pixels
een involved in
late matching to
neural network
simulated linear
s in classified
Cohonen's SOM
WORKS
self-organizing
| learning of the
Fhe topology of
exagonal. The
ht) vector of
(1)
ilar coordinate
the first is the
' second one is
gorithm starts
ust be found by
(3)
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
The second step of the learning algorithm is a weight update for
epoch /+1 in a simple way:
m(+1)=m,()+h,(0)[x-m,(0)] » (4)
where / means the epoch, and the coefficient function is defined
as follows:
a(t) if ieN,(t) (5)
hr)
a ) 0 otherwise
This formula contains the speciality of the Kohonen model,
namely the time dependent consideration of the neurons'
neighborhood in form of N,(f). The neighborhood can be
interpreted as concentric squares around the winner in case of a
rectangular neuron lattice. If the neighboring neurons are in the
acceptable region, their weights will be updated by a factor of
a(t). For this function the limits are known: it must be
between 0 and 1, mostly it decreases monotonically.
3. THE SONG MODEL
As the author of this paper applied the Kohonen type neuron
model for road junctions, great anomalies were recognized. This
was caused by the regularity of the grid structure, which was
quite far from the graph-like road crossings.
The basic idea of the extension of the SOM model was to
integrate the graph structure with its flexibility in form. The
newly named Se/f Organizing Neuron Graph (shortly SONG)
model has an undirected acyclic graph, where the computational
elements are located in the graph nodes. The edges of the graph
ensures the connections between the participating neurons. The
learning algorithm of the SONG technique has kept the
Kohonen rule for winner selection and weight update, but the
neighborhood had to be formulated in a different way.
The connections (edges) of the graph are mostly given by the
adjacency, commonly in form of a matrix. The elements of the
matrix are defined after Equation 6:
_ [1 if node i and j are connected (6)
" [|0 otherwise
This matrix is theoretically a symmetric square matrix of q X q.
The nonzero matrix elements represent the direct relations
between two neurons. The term adjacency can be extended in
order to express more information than a binary "direct
neighbor" or not. Therefore the graph algorithms are
implemented, which derive the generalized adjacency matrix
A“, where k x n . This generalized adjacency matrix has only
zero elements in the main diagonal (there are no loops), all the
other values give the number of the intermediate edges between
two nodes. The most simple generalization algorithms are the
matrix power technique (Henley 1973; Swamy 1981) and the
Floyd-Warshall method (Warshall 1962; Reinhard 2002).
The generalized adjacency matrix makes the modification of the
learning coefficient formula for graphs possible:
alt) ifA* «d(r) (0)
0 otherwise
In the above expression the condition is evaluated for the row of
the winner neuron of the matrix.
As it was mentioned regarding the Kohonen model, there is an
ordering and a tuning phase. These two computational blocks
are inherited in the SONG technique, too. Because the graph
adjacency is invariant during the run (the connections of the
neurons are fixed), the adjacency matrix can be created prior the
iterations (Barsi 2003a).
The calculations have been accelerated applying buffering
around the neurons. Such buffers are to be interpreted as
boundaries for the potential data points, so the distance
computations and therefore the whole processor load can be
limited. Figure 1 illustrates the generalized adjacency in gray
shades and the buffers in case of a letter figured graph.
154
05}
Figure 1. Neuron graph with buffers, where the neurons are
colored by the generalized adjacency values
The described SONG model is based on the adjacency
information. During the developments an other type has been
created, which is based on the distance between the neurons. In
the distance SONG model the generalized distance matrix D" is
used instead of A" , but this second algorithm is significantly
slower because of the lack of the invariance of the distance
matrix: the neuron weights (positions) are permanently
changing in every iteration steps, so the distance must be
modified (Barsi 2004). This results that the distance SONG
method can only be applied in refinement of the positions of the
adjacency SONG algorithm.
4. RESULTS
The application of the SONG technique for image analyzing
tasks is discussed in three types of examples. The first example
is a template matching use, where the previously given fiducial
structure has to be matched. The complexity of the graphs is
shown in the second group of experiments, where a given
building structure has to be found and the structure of a
complex building has to describe in form of a neuron graph. The
last tests implement a running window-type application of the
SONG algorithm: segments of a road network are detected by a
kernel graph.