Distance metric learning for content identification

Abstract

This thesis considers two distance metric learning algorithms for a fingerprinting system, which identifies a query content by finding the fingerprint in the database (DB) that measures the shortest distance to the query fingerprint. The algorithms produce a metric which improves the identification performance of the fingerprinting system for a given fingerprint DB and a set of distortions which may not have been considered when initially designing the fingerprint. In the algorithms, with a given training data set consisting of the fingerprint of the distorted and the original contents, a distance metric is learned. The first distance metric learning algorithm learns a distance metric which is parameterized by a linear projection matrix. For a given training data set consisting of original and distorted fingerprints, a distance metric equivalent to the $\It{l_p}$ norm of the difference between two linearly projected fingerprints is learned by minimizing the false positive rate (probability of perceptually dissimilar content to be identified as being similar) for a given false negative rate (probability of perceptually similar content to be identified as being dissimilar). The second distance metric learning algorithm learns a boosted distance metric which is obtained by combining various base distance metrics where the base distance metrics and the combining rule are determined by the distance metric learning algorithm. The boosted distance metric can obtain the trade-off relation between identification performance and execution time of fingerprinting system. In our experiment, the distance metric learning is applied to both audio and video fingerprinting systems. It is experimentally shown that the distance metrics learned by both distance metric learning algorithms can improve the identification performance of the fingerprinting system. It is also shown that the distance metric learned by the first distance metric learning algorithm performed better than t...