BibTeX entry

@PHDTHESIS{200305George_Gagaudakis,
  AUTHOR={George Gagaudakis},
  TITLE={Content based image retrieval using histograms},
  SCHOOL={University of Wales},
  MONTH=May,
  YEAR=2003,
  URL={http://www.bmva.org/theses/2003/2003-gagaudakis.pdf},
}

Abstract

Recent technological advances in hardware have resulted in a rapid increase in the use of pictorial data. Successful utilisation of these vast volumes of information requires effective ways of retrieval. Content-Based Image Retrieval (CBIR) was proposed as a solution to this, utilising different image aspects directly for the retrieval of relevant images from a database without the need of human involvement. These aspects are typically low-level features (colour, texture, shape etc.) extracted automatically from images. The major part of this thesis is related to methods for extracting image features that result in histograms. A number of novel histogram methods based on texture, shape and spatial information (an some hybrid) of images are presented. The motivation of this approach is due to the level of robustness and effectiveness of histogram methods proposed during the early stages of CBIR. One of the main problems of histogram methods is the limitation to capture of the semantic content from images. In this thesis BEKAS is proposed as an approach to this limitation. In BEKAS histograms are utilised on local image regions and associated to semantic terms. These terms are the link to a structure that consists of a concept hierarchy and spatial templates, allowing active use of semantic (and spatial) information extracted from images to be used to query an image database. Performance evaluation is another major issue in CBIR research, as initial attempts were tested on small scale. A framework is presented in this thesis that allowed evaluation of the presented histogram methods and their combinations. Multiple datasets are used to study the methods in isolation (e.g. tolerance to transformations), and within the context of CBIR (e.g. recall and precision of retrieval). The results are encouraging indicating significant increase in performance when multiple histograms are combined.