Digital Signal Processing Reference
In-Depth Information
→
add
X
D
to temp set
i,sub
End of for
→
set
X
-
D
to
X
,
D
`
=Φ
k = k + 1
End while
For new coming data, the algorithm will update the produce number of neurons or
add new neuron with a special value.
Input:
new Gabor feature training
set
∈
.
Output:
Updated
Produce Number of Mapping Objects
Algorithm Produce•
While
X'
X'={X' ,X' ,..., X' }
⊂
,
YR
n
1
2
m
≠
φ
do
X'
in
X'
For each
i
→
X'
of
X
randomly and then
Get a consistent subset
i,sub
do the following produce:
If covered(
i,sub
X'
)
do nothing
else
if(misclassified(
i,sub
X'
))
Adjust covering space range of hyper-ball neuron
ψ
and
adjust its priority number to lower one.
else
add new neuron with higher priority number k.
end if
end if
→
add
X'
D
to temp set
i,sub
End of for
→
set
X
-
D
to
X'
,
D
`
=Φ
End while
When new samples input into constructive neural network, the priority level of
hidden neurons is strengthened or weaken. Using this method, the old information is
not destroyed (forgot) after new data learning, but can be partly fetched at any latest
priority. This produce is similar to man's learning procedure. As this updating and
learning method is only to adjust the special priority level of neurons, therefore, it can
process large data set with more effective heuristic algorithm.
For sequence video frames, we will split it into some independent images and
transform them into eigenvector in feature space using FFT. An image frame is
