Investigação de técnicas para extração de características e indexação usando redes GHSOM aplicadas à recuperação de imagens por conteúdo.

Abstract

This dissertation proposes a content-based image retrieval system with application on the World Wide Web. A literature review has been prepared, including common techniques for feature extraction, indexing, and image retrieval. The main goal of this dissertation was to evaluate the combination of several methods for feature extraction (based on color, shape, and texture), utilizing a method of indexing based on Growing Hierarchical Self-Organizing Maps (GHSOM) with the purpose of obtaining higher accuracy. As a practical result, a system for content-based image retrieval has been developed and validated through objective and subjective experiments, in order to allow a better evaluation of the combination of implemented methods. The objective experiment used as test base a set of labeled images and had as its main goal the development of a comparative study among standalone and combined methods. On the other hand, the subjective experiment involved testing all the system and required a group of testers, who performed image retrieval experiments on unlabeled images. The purpose was to subjectively evaluate the precision of the several implemented methods. Some statistical measures have been calculated from the outputs of these experiments and promising results have been found for combinations of the features.