Adaptive Retrieval in a P2P Cloud Datacenter - Multimedia Database Retrieval: Technology and Applications - page 221

and the corresponding image files are transferred from the data owner to the query
node.
The relevance feedback is employed for re-ranking of the retrieved images, by
using the single-radial basis function (RBF) method (discussed in Chap. 2 ). The
similarity between the RBF center c

t

=[ c 1 ,..., c i ,... c P ] and the input feature vector

t , is computed by:

f v =[ f v 1 ,..., f vi ,... f vP ]

i = 1 exp − ( c i − f vi )

2

P

S ( c , f v )= (8.8)

2
i

2 ˃

where ˃ i , i = 1 ,..., P are the RBF widths. The RBF center is updated prior to the

computation of S ( c , f v ) by:

F + + ʱ N c F −

c ( t + 1 )= ( t ) − (8.9)

where F + and F − are respectively the means of the positive and negative samples,
and is the RBF center at the t -th feedback iteration,

and it is initialized by the query position, i.e.,

ʱ N is the positive constant. c ( t )

c ( t = 0 )= f q (8.10)

The relevance feedback method discussed above is referred to as multi-click RF ,
where all the top- T retrieved images that appear on the screen have to be examined
by a user. The feedback is done by the user clicking the relevant images, followed
by a feedback button. Therefore, multi-click RF requires n

+ 1 clicks, where n is the

number of relevant images appearing on the screen.
Single click RF , on the other hand, requests the user to select only one
relevant image, and this selection will be taken as a feedback sample. The system
immediately performs similarity matching using Eq. ( 8.8 ) after the single click. At
the later RF iteration, images selected on an earlier iteration will be automatically
selected as the relevant images. In this way, the single click RF considers only one
positive sample at each iteration, while the multi-click RF requires users to examine
all positive and negative images in the retrieved image set. The single-click RF can
reduce the number of clicks and the user workload, as compared to multi-click RF.
Since the negative samples are not considered in the single-click RF, the process
of updating the RBF center with Eq. ( 8.9 ) is reduced to:

F + + ʱ N c

c ( t + 1 )= ( t ) (8.11)

Figure 8.9 illustrates the retrieval results for the Corel database as used for the
experimental data in Fig. 8.5 . Fifty images of flags, balloons, bonsai, fireworks,
and ships (ten images of each category) are taken as the query images, to perform
single-click and multi-click RF for ten iterations. Figure 8.9 a compares the retrieval
precision versus iteration. The plot shows increased retrieval precision is attained