Digital Signal Processing Reference
In-Depth Information
Depending on the image patch type - for example, patches containing
higher-order structures such as axon-dendrite networks of the neurons -
a more elaborate network structure has to be found in order to avoid the
symmetry. In the following, we introduce a preprocessing method that
allows us to orient the image patches in a default way, thus enabling the
shape classification of arbitrarily oriented image patches. We will denote
trained neural classifiers employing this preprocessing as
directional
neural networks
.
The idea, similar to PCA, is to orient an image
I
along its principal
axis, where the image pixels themselves are interpreted as samples of
a two-dimensional random vector
x
I
. Hence
I
is a two-dimensional
histogram of
x
I
,andthedensityof
x
I
at the pixel (
x, y
) can be estimated
by
p
x
I
(
x, y
)
I
(
x, y
)
/T
with
T
:=
x,y
I
(
x, y
). This yields estimates
≈
for the mean
x
y
I
(
x, y
)
T
≈
μ
I
:=
x,y
E
(
x
I
)
(13.13)
and the covariance
C
I
:=
x,y
x
2
I
(
x, y
)
T
xy
−
μ
I
μ
I
.
Cov(
x
I
)
≈
(13.14)
y
2
xy
For a given image
I
,let
ρ
(
I
) be the rotated image
I
of the same size such
that the eigenvector of
C
ρ(I)
corresponding to the largest eigenvalue
is parallel to the x-axis (1
,
0). Applying the neural network training
from section 13.4 to the “normalized” training set (
ρ
(
x
λ
)
,f
(
x
λ
)) (after
adding possible reflections of the patches at the x-axis) yields the desired
directional neural network, which now is directionally selective. Any
possibly rotated input image patch is applied to the composed classifier
f
ρ
.
Figure 13.7(a) is an example of the application of the eigenvector-
based rotation. In the top row, five 15
◦
15 input patches displaying
the character
A
in various rotations and typed in various fonts are
displayed. The corresponding normalized images are given in the second
row; clearly all characters, except for the last one, were oriented such
that their main axis is parallel to the x-axis. Apparently due to aliasing
effects in the original image, in the last character the horizontal bar
of the A contributed most to the covariance, and hence could not be
rotated correctly.
×
Search WWH ::
Custom Search