CONTENT-BASED FUZZY RETRIEVAL FROM MULTIMEDIA DATABASES

Abstract

Information presentation with multimedia data like images, text, video and audio is now commonplace in most computer systems. With the advent of giga­ byte storage drives and CD-ROMs multimedia data storage is no longer a hurdle. Instead, retrieving relevant multimedia information is a challenging research problem which has still not been adequately addressed. Queries for multimedia information arc based on the contents or interpretations of the multimedia data.. Since interpretations of multimedia data like images, video and audio are not consistent among people, queries are often fuzzy and incomplete. Therefore the problem to be addressed is to develop a generic retrieval engine which can automatically relate fuzzy descriptions to its multimedia data components, support fuzzy queries, and retrieve relevant objects within all acceptable response time. In this thesis, a multimedia information retrieval system based on a fuzzy interactive and competitive (FIAC) neural network is presented as a solution to the problems stated above. A modified version of the backpropagation algorithm is used to interpret the feature measures extracted from each multimedia data object and relate them to their appropriate fuzzy sets. The retrieval engine developed based on the FIAC neural network accepts fuzzy queries and retrieves image objects. After the fuzzy query is specified, the network runs for about 20-30 cycles after which the relevant objects are picked up by the interactive mechanism and objects which are not relevant are suppressed from being retrieved by the competitive mechanism.