Visual Feature Mining for Adapting Query-by-Example over Large Image Databases
Résumé
This paper proposes a strategy for automatic selection and scheduling of visual features as search criteria that are relevant to the query-by-example on a large still-image database. This method is composed of the two main steps: the indexing and the retrieval processes. The indexing step uses K-mean clustering and association rule discovery as dimensionality reduction techniques on the descriptor values. The retrieval step is designed to order the query execution plans for speeding up the content-based retrieval over the image database. At this step, query-by-example processing is adapted in order to propose instantaneous and intermediate results that are progressively merged together with the advantage, for the users, on one hand, not to wait until the whole database has been processed by similarity search and, on the other hand, to allow them to stop the current execution of the query without losing the first partial results. We evaluate our method by comparing query execution time and result quality with the sequential search on all the low-level descriptors available in the image database. Our experiments show that we can get a result similar to the final one in less than half the time of the sequential search, which is a promising result for optimizing content-based retrieval.
Origine : Fichiers produits par l'(les) auteur(s)
Loading...