Information Technology Reference
In-Depth Information
The backward projections to the edge features expand the line features from
Layer 2 using the inverse of their excitatory forward weights. Unspecific backward
inhibition comes from
S
L
.
Layer 2: Lines.
The top layer of the binarization network is illustrated in Fig-
ure 4.15. Eight excitatory features
L
0
,L
1
,...,L
7
detect lines of different orienta-
tions. They receive 6
×
6 specific excitatory input from parallel oriented step edges.
Unspecific forward inhibition weights
S
E
with a 6
×
6 binomial kernel. The forward
projections have a bias weight of
−
0
.
05
that prevents responses to spurious lines.
Using 3
×
3 specific excitatory weights to aligned lines of same or similar orien-
tations, line cells cooperate. Competition between line features is mediated by two
inhibitory features
S
V
and
S
H
. They sum the more vertical and the more horizontal
lines, respectively, and inhibit them again. This construction restricts competition
to lines of similar orientation. Since there is no competition between horizontal and
vertical lines, crossings of two such lines can be represented without the need to
suppress one of them.
f:
←
.244(
∗F
+
∗B
) +.195
∗S
F B
-0.05
l:
←
.5(
∗E
L
+
∗E
T
+
∗E
R
) +.141
∗S
E
E
T
←
inverse of forward projections from
Lines
b:
f:
←
.244(
∗F
+
∗B
) +.195
∗S
F B
-0.05
l:
←
.5(
∗E
L
+
∗E
B
+
∗E
R
) +.141
∗S
E
E
B
b:
←
inverse of forward projections from
Lines
f:
←
.244(
∗F
+
∗B
) +.195
∗S
F B
-0.05
l:
←
.5(
∗E
T
+
∗E
R
+
∗E
B
) +.141
∗S
E
E
R
b:
←
inverse of forward projections from
Lines
f:
←
.244(
∗F
+
∗B
) +.195
∗S
F B
-0.05
l:
←
.5(
∗E
T
+
∗E
L
+
∗E
B
) +.141
∗S
E
E
L
b:
←
inverse of forward projections from
Lines
S
E
← E
T
+ E
B
+ E
R
+ E
L
Fig. 4.14.
ZIP code binarization - Layer 1 features. The image is represented in terms of
horizontal and vertical step edges.
Search WWH ::
Custom Search