Kernel-Based Adaptive Image Retrieval Methods - Multimedia Database Retrieval: Technology and Applications - page 17

To cope with the small size of training sample sets and convergence speed, new
learning methods are required for the construction of the RBF network, instead of
the direct application of traditional learning procedures. Section 2.4 introduces an
adaptive RBF network to exploit the local context defied by query sessions, and aids
in improving retrieval accuracy. This section follows by the optimization of network
parameters by the gradient-descent-based learning procedure, then introducing
fuzzy RBF network which offers a soft-decision choice to the users.
Section 2.5 establishes the fusion of content and context information, by the
application of Bayesian theory. The content component is gathered from a short-
term relevance feedback (STRF), which is the estimation of the likelihood of a
specific query model. The context information is obtained by a long-term relevance
feedback (LTRF), representing a user history or the apriori information.

2.2 Kernel Methods in Adaptive Image Retrieval

2.2.1 Adaptive Retrieval Framework

The most important part in the adaptive process is the analysis of the role of the
user in perceiving image similarity according to preferred image selections. This is
implemented by a mapping function, f q :

P

R ₒ R , which is given by:

y q = f q ( x ) (2.1)

t

where x =[ x 1 ,..., x P ] is called a feature vector in a P -dimensional Euclidean space

P , corresponding to an image in the database. The main procedure is to obtain
the mapping function f q (for the query class q ) from a small set of training images,
T = { (

R

, where the class label l i can be in binary or
non-binary form. In the binary form, the training samples contains a set of positive
samples,

x 1 , l 1 ) , ( x 2 , l 2 ) ,..., ( x N , l N ) }

X + and a set of negative samples, X − :

T = X + ∪X −

(2.2)

X + = x i | 1 , l i = i = 1 ,..., N p (2.3)

X − = x j | 0 , l j = j = 1 ,..., N n (2.4)

where N p and N n are the numbers of positive and negative samples, respectively.
The adaptive process for constructing the mapping function for retrieval is
summarized in Table 2.1 .