Model-based decoding metrics for content identification

In this paper, decoding metrics are designed for statistical fingerprint-based content identification. A fairly general class of structured codes is considered, and a statistical model for the resulting fingerprints and their degraded versions (following miscellaneous content distortions) is proposed and validated. The Maximum-Likelihood fingerprint decoder derived from this model is shown to considerably improve upon previous decoders based on the Hamming metric. A GLRT test is also proposed and evaluated to deal with unknown distortion channels.