Evaluating Retrieval Effectiveness of Descriptors for Searching in Large Image Databases

Authors

  • Petrina A. S. Kimura Federal University of Amazonas
  • João M. B. Cavalcanti Federal University of Amazonas
  • Patricia C. Saraiva Federal University of Amazonas
  • Ricardo Da S. Torres Instituto de Computacao, UNICAMP
  • Marcos A. Gonçalves Federal University of Minas Gerais

Keywords:

content-based image retrieval, experimental evaluation, image databases

Abstract

This article presents an evaluation of image descriptors for searching in large image databases. Several image descriptors proposed in the literature achieve high precision levels when experimented in small (less than 20,000 images) and well-behaved image databases. Our assumption is that retrieval effectiveness may be  strongly affected by variations in size, quality, and diversity of the images in the database. In order to verify whether an image descriptor maintains its retrieval effectiveness in large databases, experiments were carried out using several image descriptors and three image collections, including one with over 100,000 images collected from the Web. The results obtained show that in general the retrieval effectiveness of the different descriptors varies little in small image collections whereas in large image collections they differ significantly. Among the descriptors used in the experiments, there are two proposed by us for being used in large and heterogeneous image databases. The proposed descriptors outperform significantly the other descriptors used as baselines in the Web collection. These results give us a better understanding about the features and the strategies that should be followed to construct descriptors for practical search tasks in large image databases.

Downloads

Download data is not yet available.

Downloads

Published

2011-09-13

Issue

Section

SBBD Articles