Tracing Images in Large Databases by Comparison of Wavelet Fingerprints

This paper presents an algorithm that extracts the pictorial information of an image into a very compact wavelet fingerprint while conserving the most significant global and local visual characteristics of the image. A query algorithm is described that compares a fingerprint of a query image with all potential target fingerprints from a small remote database in terms of their similarity. The similarity measure itself is based on a multiresolution representation of the fingerprints attempting to simulate the human ability to abstract information at different levels of detail. The robustness of this technique allows query specifications with a significant amount of distortions (e.g. sketches). The remarkable speed of the query algorithm allows progressive query refinements even for large databases (up to 100.000 images) on commodity platforms. Empirical as well as analytical query results are given.