Digital Signal Processing Reference
In-Depth Information
(
a
)
(
b
)
Figure 12.6
Typical spreading of a Kohonen chain over 16 patches of a psoriasis image. Only
patches which can already be represented adequately are indicated by a
corresponding neuron. After convergence of the network, all patches can be
described by the weight vectors
w
(j)
of the Kohonen chain.
learning rate
η
(
t
) is decreased. Thus the one-dimensional Kohonen chain
quickly spreads over the image data at the beginning of the simulation
and subsequently is fine-tuned for an accurate representation of the cor-
responding patch cluster in the end.
This procedure is illustrated in fig 12.6, where the spreading of
a Kohonen chain over 16 patches of a psoriasis image is displayed:
Note that for the sake of clarity, only patches which can already be
Search WWH ::
Custom Search