Merging Image Features by Self-organizing Maps in Content-based Image Retrieval

By Christian Breiteneder, Dieter Merkl, and Horst Eidenberger

Abstract

This paper shows how merging by linear combination of weighted distance values can be performed. A weighting algorithm is presented for the automatic computation of suitable weights for image features. These features are arranged in query models. The algorithm bases on a self organizing map which describes the "natural" clusters within an image database. We showed that using the contribution of a feature to the cluster structure as weight improves the ordering of query results. The implementation uses IBM QBIC system as kernel and runs under LINUX.

Reference

Proceedings of European Conference on Electronic Imaging and the Visual Arts, Berlin, Germany, 1999