Kernel-Based Adaptive Image Retrieval Methods - Multimedia Database Retrieval: Technology and Applications - page 31

Fig. 2.2 Top 16 retrieved images obtained by the Yacht query, using the Corel Database, ( a )before
RF learning, ( b ) after RF learning with the RBF method, ( c )MAM,and( d )OPT-RF

Previously, Sect. 2.3 introduced a nonlinear input-output mapping function based
on a single-RBF model. As discussed by most other works [ 15 - 17 ], this has been
concerned with global modeling, in which a query image is described by one
model, which is then associated with only a particular location in the input space.
Furthermore, the similarity function is based on a single metric. This combination
gives rise to a single model function f , which cannot fully exploit the local data

information. This section introduces a mixture of Gaussian models for adaptive
retrieval that enables the learning system to take advantage of the information from
multiple sub-classes. The learning system utilizes a highly local characterization of
image relevancy in the form of a superposition of different local models, as

( x )

∑ i f i (

x ) ,

to obtain the input-output mapping function.
The learning methods for constructing the RBF network include the adaptive
RBF method [ 61 ], gradient-descent method [ 40 ], and fuzzy RBF method [ 39 ].