Abstract
This thesis presents a novel paradigm known as the 'Pictorial Dictionary' for the content-based image retrieval problem, which has been proposed in order to avoid the limitations inherent with conventional content-based image retrieval approaches. With conventional content-based retrieval systems, the user often has two methods by which a query can be formulated, the 'direct query' method and the 'query by example' method. Both of these methods limit the user in their ability to easily formulate queries. The Pictorial Dictionary scheme allows users to easily formulate queries through a high-level textual interface. The objects in the dictionary are used to provide an association between low-level features and high-level semantics. In order to research this paradigm, both low-level and high-level approaches were investigated. The techniques explored included the pre-processing techniques and feature matching that were required to complete the experimentation. Colour matching methods were explored, leading to the proposal of a metric which was suitable for matching in the presence of varying illumination. A suitable segmentation method was presented which allowed the low-level retrieval problem to be formulated as attributed relational graphs. Three graph matching methods are presented and evaluated in the context of this paradigm. Natural language processing techniques are also explored within the context of our pictorial dictionary. An algorithm for the measurement of semantic distances was proposed and qualitatively evaluated in the pictorial dictionary testbed. The computation of semantic distance allows items in our dictionary to be indexed and related through semantic content while avoiding common problems associated with vocabulary and human annotations.