nextuppreviouscontents
Next:Interactive Model Recognition from Up:Computer Vision Previous:Autonomous Physics-Based Colour Learning

An Improved Appearance-based Approach to Image Retrieval and Classification

F. Beyrouti, F.P. Ferrie

A content-based image retrieval system operates by matching indices that are based on the contents or structure of an image as opposed to annotations included as part of the database representation. As part of the Digital Library Project, a collaborative effort among researchers from Concordia University, McGill University, Université de Montréal, and Université du Québec à Montréal, the goal of this work is to investigate how content-based retrieval can be integrated into existing standards for digital libraries. The idea is to complement traditional annotations with indices generated from scene content in a query by example context. We are currently investigating a two-step hierarchical approach that first attempts to assign a query image to a restricted set of classes within the database, and then returns the best matches to each of the selected classes. Each class has associated with it specialized pre-filtering intended to enhance the selectivity of its associated pattern classifier. Since appearance-based methods are used for classification, this pre-filtering also serves to reduce the sensitivity of the classifier to features that are not used for indexing. The database organization also helps to speed up retrieval since only a subset of the database need be considered once the applicable classes have been determined. Preliminary experiments indicate that this approach appears to be well-suited for query by example applications. We are currently investigating the relationship between different pre-filtering approaches and scene categories. Trials will then be performed on larger scale image databases and our work integrated with that of the rest of the consortium.
 


Annual Report

Fri Nov 26 23:00:32 GMT 1999