Information Technology Reference
In-Depth Information
and the complexity of the task increases with the number of different classes,
however, the use of NN algorithms can correctly resolved effectively circumvent
the problem. The error is calculated by the equation:
error
t
=
S
t
(
i
)[
|
h
t
(
x
i
)
=
y
i
|
]
(11)
i
:
h
i
(
x
i
)=
y
i
with
h
t
:
X
Y
an hypothesis and where
TR
t
is the subset of training subset
and the
TE
t
is the test subset. The synaptic coecients are updated using the
following equation:
→
β
t
if H
t
(
x
i
)=
y
i
1
else
w
t
+1
(
i
)=
w
t
(
i
)
∗
(12)
Where
t
is the iteration number,
B
t
composite error and standard composite
hypothesis
H
t
.
Fig. 5.
M-SVM classifier
In our approach we replace each weak classifier by SVM. After
T
k
classifiers are
generated for each
D
k
, the final ensemble of SVMs is obtained by the weighted
majority of all composite SVMs:
K
log
1
β
t
H
final
(
x
)=
arg max
(13)
y
k
=1
t
:
h
t
(
x
)=
y
∈
Y
5 Visual Feature Extraction
We use a set of different visual descriptors at various granularities for each frame,
rid of the static background, of the video shots. The relative performance of the
Search WWH ::
Custom Search