Database Reference
In-Depth Information
Table 2.15
Comparison of learning methods
Method
Learning criterion
STRF
LT R F
NN-CBIR
Adaptive image retrieval method, using
nearest neighbor
(NN)
criterion, where L1-norm is employed as the distance function and
Eq. (
2.8
) is employed for a query modification
×
SVMAL-
CBIR
Adaptive image retrieval method using
support vector machine
active learning
(SVMAL)
×
NN-CLBIR
Collaborative Bayesian Image Retrieval (CLBIR), using Eq. (
2.84
)
for estimation of
p
(
x
q
|
c
)
and Eq. (
2.87
) for estimation of
P
(
c
|
I
)
SVMAL-
CLBIR
Collaborative Bayesian Image Retrieval (CLBIR), using Eq. (
2.85
)
for estimation of
p
(
x
q
|
c
)
and Eq. (
2.87
) for estimation of
P
(
c
|
I
)
the query-by-example retrieval paradigm, the average of the features of the two
images is used as the feature of an exemplar image.
To facilitate the subsequent elaboration, the query subsets which consist of
the first five queries, the sixth through the eighth, and the ninth and the tenth
in each class, are denoted
T
A
,
T
B
,
1
,
and
T
B
,
2
, where
|
T
A
|
=
1
,
000,
|
T
B
,
1
|
=
400,
and
600. Such a query set selection guarantees that the system trained
using the LTRF will be tested based on previously unseen samples.
T
A
was used
when there is no accumulated high-level knowledge, i.e. before LTRF happens.
In such a case, only STRF is involved, and the “nearest neighbor collaborative
Bayesian image retrieval” (NN-CLBIR) and the “support vector machine active
learning collaborative Bayesian image retrieval” (SVMAL-CLBIR) are essentially
the same as the NN-CBIR and SVMAL-CBIR because the
apriori
distribution
of the candidate images is uniform. After the initial LTRF, the CLBIR systems
are expected to present better performance in general thanks to the accumulated
knowledge, while the STRF still improves the results with respect to each specific
query.
T
B
,
1
∪
|
T
B
,
2
|
=
T
B
,
2
, comprising 1,000 images, was used to verify the improvement
after the initial LTRF. During the operation of the CLBIR systems, the new retrieval
results after the initial LTRF are gradually accumulated, and a second LTRF can
be carried out upon a certain point. The retrieval results corresponding to
T
B
,
1
were
used to perform an incremental update of the system, i.e. the second LTRF, after
which the performance was evaluated using
T
B
,
2
.
To capture various visual properties of the images, three types of low-level
descriptors are selected, including global color histogram in Hue-Saturation-Value
(HSV) space, color layout in YCbCr space [
92
], as well as Gabor wavelet [
91
].
Shown in Fig.
2.8
a is the comparison between NN-CBIR and NN-CLBIR in
terms of the average precision
Pr
as a function of the number of iterations of
STRF. The precision is given by
Pr
N
R
, where
N
C
and
N
R
are the numbers
of relevant images and retrieved images, respectively. The precision is measured
in the top
N
R
=
=
N
C
/
48 in this case. Using the query set
T
B
,
1
, the improvement due to
LTRF based on past retrieval results with respect to the query set
T
A
is obvious, and
the effect of STRF can also be observed. After the second LTRF, the performance
of NN-CLBIR using query set
T
B
,
2
is further enhanced due to more accumulated
Search WWH ::
Custom Search