Multimedia copy detection

Asmultimedia-sharing websites are becoming increasingly popular, content providers get more concerned about the illegal distribution of their copyrighted contents. The recent content-based multimedia fingerprinting technology has evolved as an important tool for automatically detecting illegal copies of audio, image, and video signals. Multimedia fingerprints are signatures that are extracted from an audio, image, or video signal as a compact identifier of the signal. Therefore fingerprints should have enough discriminating ability to identify a multimedia object among others. At the same time they should be robust to modifications a multimedia signal might be subjected, such as compression, cropping, format change, scaling, and other signal processing operations. Robustness requires the fingerprints of a signal to only depend on the signals perceptual content and not on its format, size, quality, etc. This thesis proposes copy detection systems for audio and video signals and addresses the robustness as well as the discrimination ability of these systems. We first address audio fingerprinting and propose an algorithm that can detect small snippets of audio signals. Simulation results show that, the extracted fingerprints are robust to audio modifications including pitch shift and tempo change. For severe modifications that existing algorithms have poor detection rates (around 20%), our proposed algorithm yields detection rates above 80%. We then address video fingerprinting and propose an algorithm that extracts robust and discriminant binary fingerprints. Simulation results show that the proposed algorithm is faster and more accurate than the state-of-the-art with a high true positive rate of over 97% and a low false positive rate below 1%. Another challenge in multimedia fingerprinting is fingerprint retrieval, i.e. searching a huge fingerprint database (millions of fingerprints), for an accurate match for a query fingerprint in a fast fashion. We propose a fast and accurate Nearest Neighbour (NN) search algorithm for binary fingerprints (Hamming space). Tested on a very large database of 80 million images, we showed that the proposed algorithm is about 3 times faster than the state-of-the-art while at the same time it is 10 times more accurate.Applied Science, Faculty ofElectrical and Computer Engineering, Department ofGraduat