Kernel-Based Adaptive Image Retrieval Methods - Multimedia Database Retrieval: Technology and Applications - page 50

the system is capable of integrating the useful information predicted using the
context component and that learned using the content component. The a posteriori
probability evaluated by the system is used to rank the images in the database.

2.5.2 Content-Based Likelihood Evaluation

in Short-Term Learning

The visual content model of a certain semantic class, e.g. c , is the parametric form
of the distribution of the visual features of that class. The parameters of the model
are adapted to a given set of training data of class c through a supervised learning
procedure. A visual content model plays the role of evaluating the likelihood of a
visual feature with respect to a certain class. The support vector machine (SVM)
is selected as the key component of the content model to evaluate the likelihood.
L1 norm is also employed in addition to SVM for calculating the likelihood using
the content model. At the same time, it should be noted that the formulation of the
Bayesian framework requires that the output of the visual content model comply
with the definition of a PDF. To this end, the exponential function is employed, i.e.
h , to convert the discriminant function of SVM into a PDF. The

selection of the above exponential function is based on the following consideration.
First, it is monotonically increasing, resulting in the preservation of the physical
interpretation of the algebraic distance between a sample and the decision boundary.
Second, it is positive. Since the total integral of a function must be equal to
unity, appropriate normalization is necessary. Finally, representing the discriminant
function of SVM corresponding to the c -th class as f c (

( s )= exp ( s ) , s ∈ R

and substituting it for the
variable s in the exponential function followed by normalization, we obtain

x )

1
A exp

p ( x | c )= ( f c ( x )) (2.83)

exp

where A = ( f c ( x )) dx .

2.5.2.1 Using the Nearest Neighbor (NN) Method

The nearest neighbor (NN) method returns the top K images on the list, which
is ranked based on the similarity measure between the feature of the query and
that of each of the candidate images, where K

C . The L1-norm is used as the
distance function for the NN method. In adaptive retrieval, the query is refined using
the method of query point movement [i.e., Eq. ( 2.8 )]. To calculate the likelihood,
the exponential function in Eq. ( 2.83 ) converts the L1-Norm into a similarity
function, i.e.

A exp − x q −

x c

1
A exp

1

p ( x q | c )= ( f c ( x q )) = (2.84)