DCT-Based Videoprinting on Saliency-Consistent Regions for Detecting Video Copies with Text Insertion

Ideal video fingerprinting should be robust to various practical distortions. Conventional fingerprinting mainly copes with natural distortions (brightness change, resolution reduction, etc.), while always gives poor performance in case of text insertion. One alterative way is to apply a weighting scheme based on the probability of text insertion for feature similarity calculation. However, the weights must be learned with labeled samples. In this paper, we propose a method that first addresses valid regions where the saliency values keep consistent between the query and original frames, namely saliency-consistent regions. Other regions, probably the inserted ones, are discarded. Then a DCT-based hamming distance is calculated on those saliency-consistent regions. Besides, the saliency-based distance is also considered and a further weighted linear distance is evaluated. The proposed algorithm is tested on the MPEG-7 video fingerprint dataset, achieving a false rate of 0.7% in case of text insertion and 0.32% in average for other 8 distortions.