Distance combination for content identification system

This paper considers an algorithm which produces a distance function by combining distance functions for a finger-printing system, which identifies a query content by matching its fingerprint to the database (DB) fingerprint. To match finger-prints, recent audio and video fingerprinting systems commonly use a simple distance metric such as l₂ distance, and output the information of DB fingerprint that measure the shortest distance to the fingerprint of the query. This paper considers the weighted sum among various combining methods, and the weights are determined by the learning process with a given set of training data, which consists of the fingerprint of the distorted and the original contents. By solving an optimization problem which reduces the fingerprint distance between similar contents and increases the fingerprint distance between dissimilar contents, weights are determined. In our experiments, the proposed algorithm is applied to a video fingerprinting system, and the experimental results shows that combined distance outperforms the conventional l₂ distance and that out algorithm to determine weights is reasonable.