Database Reference
In-Depth Information
Table 2.14
Average
precisions,
Pr
, compared at
four settings of top matches
(
N
c
), obtained by retrieving
59 queries, using the Corel
Database
Average precision (%),
Pr
(
N
c
)
N
c
Method
t
=
0
t
=
1
t
=
2
t
=
3
10
ARBFN
55.93
+
32.03
+
42.03
+
43.56
Single-RBF
55.93
+
27.97
+
39.15
+
42.03
MAM
55.93
+
17.12
+
19.32
+
19.66
+
+
+
OPT-RF
55.93
24.07
30.51
32.37
16
ARBFN
47.67
+
30.83
+
39.30
+
41.21
+
+
+
Single-RBF
47.67
26.48
34.64
38.45
+
13.88
+
16.00
+
16.21
MAM
47.67
OPT-RF
47.67
+
20.97
+
23.83
+
25.00
25
ARBFN
39.93
+
26.44
+
30.44
+
31.19
Single-RBF
39.93
+
21.36
+
26.58
+
28.14
MAM
39.93
+
11.46
+
12.47
+
12.07
OPT-RF
39.93
+
17.02
+
19.73
+
20.00
50
ARBFN
30.03
+
19.08
+
20.58
+
20.75
Single-RBF
30.03
+
15.29
+
17.76
+
18.44
MAM
30.03
+
8.24
+
8.31
+
8.17
12.51
Interactive results are quoted relative to the
Pr
observed with
the initial retrieval
OPT-RF
30.03
+
11.86
+
12.17
+
2.5
Bayesian Method for Fusion of Content and Context
in Adaptive Retrieval
Adaptive retrieval method can be implemented to integrate visual content and
contextual information through relevance feedback [
47
,
48
]. Contextual information
refers to the statistical correlation across multiple images. In this section, a Bayesian
framework is developed for fusion of content and context components. Specifically,
the visual content analysis is associated with the likelihood evaluation, whereas the
contextual information is represented by the
apriori
probability, learned through a
maximum entropy algorithm.
2.5.1
Fusion of Content and Context
Let
C
represent the set of class labels and
C
, where
C
is the number of
classes. The class label of a particular image in a database is denoted
c
, where
c
=
{
1
,
2
,...,
C
}
C
.
Basedonthe
maximum a posteriori
probability (MAP) criterion which minimizes
the classification error, the true class label is estimated with:
∈
c
=
arg max
c
P
(
c
|
x
,
I
)
(2.80)
∈
C
Search WWH ::
Custom Search