Distance metric learning for content identification
Dalwon Jang, Chang D. Yoo, and Ton Kalker
This research considers an algorithm to determine a boosted distance metric for a video fingerprinting system, which identifies a query video by finding the fingerprint in a database (DB) that measures the shortest distance to the fingerprint of the query. In this paper, the fingerprinting task of identifying a video is cast into a binary classification task of matching and nonmatching fingerprint pairs: a matching fingerprint pair should be declared as being extracted from identical content, and a nonmatching fingerprint pair should be declared as being extracted from different contents. In this study, the boosted distance metric is represented as a weighted sum of base distance metrics, and for a given training data set of matching and nonmatching fingerprint pairs, these base metrics and their weight are determined iteratively such that the subsequent decision stump classifier determined by a base metric and threshold is designed in favor of those fingerprint pairs misclassified by a previous classifier. By varying the number of base distance metrics used in constructing the boosted distance metric, a trade-off between the distance computing time in computing the distance and identification performance can be made. Experimental results show that the boosted distance metric can improve the identification performance of the fingerprinting system over the l2 distance metric.
1. Dalwon Jang, Chang D. Yoo, and Ton Kalker, "Distance metric learning for content identification", IEEE Transactions on Information Forensics and Security, vol.5, no.4, pp.932-944, December 2010.
2. Dalwon Jang and Chang D. Yoo, "Fingerprint Matching Based on Distance Metric Learning", IEEE International Conference on Acoustics, Speech and Signal Processing, Taipei, Taiwan, 2009.