Accurate Data Modelling for Watermarking Applications
Mihai Mitrea, Titus B. Zaharia, Franç̧oise Prêteux, Adriana Vlad · 2004
Abstract: Having in view a watermarking application, this paper presents an original statistical approach making it possible for video sequences to be accurately modelled in the wavelet domain. The 2D-DWT (2D Discrete Wavelet Transform) is individually applied to each frame in the video sequence and the coefficients corresponding to some low-frequency sub-bands are sorted according to the decreasing order of their absolute values. In order to identify the pdf (probability density function) characterising the values taken by a certain rank in this coefficient hierarchy, the frames in the video sequence are first divided into classes with independent elements. Then, each class in the partition is investigated by successively applying the Chi-square, the Ro, the Fisher F and the Student T statistical tests. Finally, the results obtained on each class are combined so as to provide information characterising the whole video sequence. As a general result, the present study states that the Gaussian pdf is suitable in modelling a large rank interval in the DWT coefficient hierarchy. The experiments correspond to three types of transforms, namely the (2,2), (4,4) and (9,7) bi-orthogonal DWTs and to 10 video sequences of about 35000 frames each (about 25 minutes each). 1.