Database Reference
In-Depth Information
Fig. 8.12 Average precision
rate, obtained by retrieval in
the P2P network using a
simple search with and
without the application of the
occurrence weight, for
different numbers of
community neighbors
Table 8.1 Average precision rate (%) obtained by the pseudo-RF method on
the P2P network, where the bias weight
ʳ
is applied for obtaining the RBF
center [cf. Eq. ( 8.14 )], and the parameter
ʲ
is applied for the SOTM [cf.
Eq. ( 8.13 )]
With
ʲ
Bias weight
Iter. 0
Iter. 1
Iter. 2
Iter. 3
Iter. 4
Iter. 5
ʳ =
0
44.75
57.15
68.30
77.45
79.50
82.05
ʳ =
0
.
2
44.75
58.15
71.35
77.90
80.85
83.25
ʳ =
0
.
2
×
44.75
57.10
68.60
76.15
79.20
82.30
categories which each peer falls under followed a normal distribution, with mean
μ cat =
10 and standard deviation
˃ cat =
2. The number of images per category was
also normally distributed, with mean
5.
Firstly, the retrieval system was conducted by a simple search strategy, where
image ranking was done by an occurrence weighting scheme [cf. Eq. ( 8.12 )].
Figure 8.12 shows the statistical analysis of the size of social network with respect
to retrieval precision. It was observed that the retrieval precision steadily increased
against the size of the community neighborhood. Such characteristics serve as the
foundation of the current P2P retrieval system.
Next, the retrieval system utilized a pseudo-RF approach, where SOTM was
employed for pseudo labeling and the single RBF method was employed for
similarity measurement. Since the pseudo labeling can cause an error in the
classification of image relevancy, the bias weight parameter
μ image =
50, and standard deviation
˃ image =
was used for updating
the RBF center [cf. Eq. ( 8.14 )]. Table 8.1 shows the average precision as a function
of feedback iterations, when the system utilized
ʳ
0 and 0.2. Regardless of the
setting of the bias weight, the retrieval performance significantly improved from
44.75 to 83.25 % at the fifth iteration. At each iteration, the system with
ʳ =
ʳ =
0
.
2
provided better retrieval accuracy than that of the system with
0. This system
adaptively improved retrieval accuracy without user interaction. Figure 8.13 shows
ʳ =
 
Search WWH ::




Custom Search