centroids in oth-
le demonstrating
in Fig. l. Since
riant, a normal-
pensate this ef-
ting the longest
distances in the
tio between the
input image and
1e object model,
e divided by this
each polygon in
viding it by the
g ; (b)Its poly-
olygon 2, local
onal features:
put image or an
ibeling process.
polygon in the
nd each polygon
umn index. An
f
112
ANZ
\ AS À
, 2 4
1
2
0 3
nc 5
\ 16 $
1 8
BN 10
12
Fig.2. Row-column assignment for
surface correspondence establishment
B. Cj for Surface Matching
At the stage of surface correspondence establish-
ment. Cj; is expressed as follows :
Cup = W1 X F;,M + W 9 X F(;,M,)
+ Wy X Fy sduym) (13)
where I, represents the (normalized) area of the zth
polygon in the input image, M, the area of the yth
polygon in the object model, dy, the (normalized )
distance between the centroids of the ith and the jth
polygons in the input image, and dum, the distance
between the centroids of the kth and the Ith polygons
in the object model. The values of W ;s reflect the im-
portance of each term. They can be adjusted as long
as the sum equals to 1. For the symmetric terms, the
associated weights should be set equal (e.g. , W,;=
W,). The weight of the relational feature (W3) in
more important than the other two and is thus set
with higher value.
C. Similarity Measure from Surface Matching
After the states of the Hopfield net for surface
matching are stabilized, we can count the number of
active neurons in the network and use it to measure
the degree of match (or similarity) between the ob-
ject model and the input image. The procedure con-
sists of the following four steps.
Step 1: Initialize both. row. match and column...
match to be 0.
Step 2; Count the number of l's in each row. If
there is nol in a row, skip to the next row and leave
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
row _ match unchanged. If there is only one lina
row , add 1 to row match. If there are n 1s (421) in
a row, then add 1/n to row match. Repeat this for
all the rows.
Step 3: Do the same ealculation for all the columns
and updaté column. match.
Step 4; Pick up the larger one from row _ match
and column .maích , divide it by the number of rows,
and take the result as the similarity measure.
Given an input image and a large number of object
models (which is usually the case in a multiple-view
approach), the degree of match between the input
image and different object models can be derived by
comparing the input image and each of the object
models in the final state of the Hopfield net. Ideally,
there should be at most one active neuron in each row
or column. However, due to the influence of the first
term in the right-hand side of (1),'it is possible to
have more than one candidate in the same row or col-
umn in the final state of the network. As far as
matching is concerned, this situation should be con-
sidered as unfavorable and hence decreases the degree
of match. This is the reason why 1/» is added to the
degree of match (row or column) instead of ] when
there are 4 1s simultaneously existing in the same
row or column. When there is no 1 in a row (or col-
umn), it means a surface in the input image (or ob-
ject model) does not have a corresponding surface in
the object model (or input image). This does not
contribute to the degree of match between the input
image and the object model and thus the degree of
match is left unchanged. For a model-based 3-D ob-
ject recognition system using multiple-view approach ,
a set of 2-D object models are in the model database.
To derive their degrees of match with the input im-
age in the Hopfield net, we associate the input image
with row indexes and object models column indexes.
This arrangement allows us to compare all the object
models simultaneously with the input image, provid-
ed that the dimension of the neuron array is large e-
nough. This also explains why the number of rows is
used as the denominator in the derivation of similarity
measure.
D . Discussion
The proposed Hopfield net for surface matching
has a flexible structure and is able to solve the surface
correspondence problem even if the numbers of poly-
gons in the input image and the object model are dif-
ferent. in other words, the two-dimensional neuron
array may have different numbers of rows and
columns and an inexact matching [18] can be per-
formed in this net. Furthermore, based on the out-
puts of the neurons in the network , a similarity mea-
1013