Database Reference
In-Depth Information
Table 2.3
Summary of the single-RBF based relevance feedback algorithm
Input:
Query vector
=
x
q
The training samples
N
i
=
{
x
i
}
=
1
Output:
The final retrieval set, containing
k
-relevant samples
=
S
k
(
x
q
)
Initialization:
RBF center
z
c
≡
x
q
Repeat:
User labels training samples, i.e.,
l
i
,
i
=
1
,...,
N
,
l
i
∈{
0
,
1
}
+
ʱ
R
x
−
z
ol
c
−
ʱ
N
x
−
z
ol
c
Calculate RBF center:
z
new
c
z
old
c
=
Calculate RBF widths:
ma
m
x
mi
−
z
i
,
i
=
˃
i
=
ʷ
,...,
P
1
P
i
=
1
G
(
x
ni
,
z
i
)
Calculate
f
q
(
x
n
)=
∑
,for
n
=
,
,...,
T
, and obtain
1
2
S
k
(
x
q
)=
x
|
f
q
(
x
)
≥
f
q
(
x
k
)
where
S
k
(
x
q
)
is the set of nearest neighbors and
x
k
is the
k
-th nearest neighbor
of
x
q
.
N
i
{
x
i
}
ₐ
S
k
(
x
q
)
=
1
Until:
User is satisfied with the retrieval result.
Table 2.4
Database and feature extraction methods
Item
Description
Brodatz
texture
The database contains 1,856 texture images divided into 116 classes. Every
class has 16 images
database
Laplacian
mixture model
(LMM) [
35
]
The images are decomposed to three levels using Daubechies wavelet filters
(db4). The wavelet coefficients in each of high-frequency subbands are
modeled as a mixture of two Laplacians. The parameters of the model are
used as the features. The feature set composes of (1) the mean and standard
deviation of the wavelet coefficients in the approximation subbands and
(2) the variances of the two Laplacians in each of the nine high-frequency
subbands. This results in 20-dimensional feature vector
In the second experiment, the adaptive retrieval methods are applied in photo-
graph collection. Table
2.7
gives details of the database and the multiple types of
visual descriptors, including color, texture, and shape. Table
2.8
gives details of the
methods being compared. The average precision rates and CPU times required are
summarized in Table
2.9
. Evidently, the nonlinear RBF method exhibits significant
retrieval effectiveness, while offering more flexibility than MAM and OPT-RF.
With the large, heterogeneous image collection, an initial result obtained by the
non-adaptive method had less than 50 % precision. With the application of the
RBF learning method, the performance could be improved to greater than 90 %
precision. Due to limitations in the degree of adaptability, MAM provides the
Search WWH ::
Custom Search