Uma abordagem evolutiva para recuperação de imagens da web

• Ano de defesa: 2009
• Autor(a) principal: Katia Cristina Lage dos Santos
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; UFMG
• Palavras-chave em Português: recuperação de imagens; web
• Link de acesso: http://hdl.handle.net/1843/SLSS-7WMNMH

The developments in data storage and image acquisition technologies made the creation of huge databases or image collections possible. Moreover, the Web has been providing low cost and large scale acessibility to this material. Along with these factors, a variety of profitable activities demand image-based information. Typical examples are architectural projects, engineering and fashion designs, perfumes, new car models, marketing campaings and Internet sites, among others. All these needs explain the increasing in the interest in organizing, indexing, and retrieving digital images in the last decade.This problem can be simply described as follows: given a textual or a image-based query, supplied by the user to a system which contains a image database, define the most relevant answer-set to the query.This is a challenging task due to the difficulty in extracting from the image sources the `best' information needed for their representation and indexing. This work presents a evolutionary framework for image retrieval based on the combination of multiple textual sources of evidence. It explores the Genetic Programming concepts, based on the continuous improvement of the solution quality for a given problem. Therefore, our motivation is to contribute with the development of search mecanisms capable of representing, in a more reliable manner, images in the the WWW. As a pratical result, a query made to this system will have more relevant images as result, increasing the user satisfaction and confidence on the search system.Experiments performed with a collection extracted from the Web showed that, compared to the Bayesian Model presented in ~\citep{berthier} to solve the same problem, the evolutionary framework presents a performance twice as good, under precision, recall and MAP. These are important metrics to evaluate the quality of a image set returned from a query. Besides, the computational solution developed in this work presents great flexibility since the adition of new sources of evidences will be allowed with a minimum cost, besides the possibility to use the framework to retrieve images from other collections.