An Extension of Metric Histograms for Color Based Image Retrieval

George BrindeiroAndré GeraldesDíbio Leandro Borges

Color based image retrieval is an important and challenging problem in image and object classification. Many techniques work by predefining a number of dimensions to reduce an original histogram and evaluating the image similarities using norms defined on the reduced spaces. Metric histograms, on the other hand, do not predefine this number, and explore the correlations between significant points and their neighborhoods in order to find a small number of control points to represent the histogram. It lacks though a proper way to deal with color images, since it considers only normalized gray level histograms. In this paper we propose tri­dimensional metric histograms for considering the color space. We introduce a procedure to compute a parameter to span the range of inflection points between minimum and maximum for the specific data. An extended distance metric for it is also presented. Experiments ran with a database of 2090 color images show better performance of the proposed approach than the original one.

Caso o link acima esteja inválido, faça uma busca pelo texto completo na Web: Buscar na Web

Biblioteca Digital Brasileira de Computação - Contato:
     Mantida por: