Adaptive Retrieval in a P2P Cloud Datacenter - Multimedia Database Retrieval: Technology and Applications - page 230

Fig. 8.13 Average precision
result, obtained by the
pseudo-RF method in
comparison with the
user-controlled-RF method,
on the P2P network

the comparison of retrieval performance between the user-controlled RF and the
pseudo-RF methods. It was observed that the pseudo-RF method performed very
close to that of the user-controlled RF method.
The application of the ranking bias

in updating the SOTM prototypes also
influenced the system performance. Since the output at each feedback stage was
ranked according to the RBF function, it was observed that by adding the ranking
bias

ʲ

to the construction of the prototypes as in Eq. ( 8.13 ), the retrieval precision
was improved. The result is also shown in Table 8.1 .
The inter-dependence between each individual image database can be used to
improve the retrieval precision for a centralized retrieval system, using the same
algorithm applied for a distributed retrieval system. While the centralized retrieval
system typically includes a higher order database, greater diversity is expected. In
this simulation, the number of images per category was also normally distributed,
with

ʲ

5. Comparisons between the
retrieval methods in the centralized system, the centralized system accounting for
inter-dependencies between individual databases, and the P2P system, are illustrated
in Fig. 8.14 . Here the relevance feedback was done manually. Accounting for the
overlap between relevant databases used in the distributed P2P system, as described
in Sect. 8.4.2 , it was observed that there was improvement in the retrieval precision
for the centralized retrieval system that used the occurrence weighting scheme.
Figure 8.15 compares the experimental results between the semi-automatic RF
and the user-controlled RF, applied to the P2P retrieval system. The x-axis indicates
n , the number of manual RF performed. For the semi-automatic RF, the plot
is coupled with 5

μ cat = 20, ˃ cat = 5, μ image = 50, and ˃ image =

n pseudo RF. The user-controlled RF was conducted solely
from the manual RF, whereas the semi-automatic started with the pseudo RF and
was followed by the manual RF. It can be observed that, while n is small, the
semi-automatic RF is a better method providing a significant improvement over
the user-controlled RF, where both methods use the same amount of manual RF.

−